• The Australian Competition & Consumer Commission (ACCC) has bricked thousands of routers
• Flashing alternative firmware like OpenWRT is an option that the ACCC has ignored
• Disposed SamKnows SK-WB8 routers can pose a security risk if they’re not first wiped with a factory reset

The Australian government’s competition regulator, the Australian Competition & Consumer Commission (ACCC), has initiated the bricking of some 4,000 routers, rolled out across the country in 2020 in order to collect information on broadband speeds.

Issued as part of the Measuring Broadband Australia (MBA) program, the routers were whiteboxes from SamKnows, a subsidiary of Cisco, and given a finite lifespan. That date was reached in June 2026, with the SamKnows SK-WB8 routers remotely disabled on June 30.

As a consequence, Australians are likely to toss the bricked routers, resulting in a sharp increase in e-waste. In doing so, they risk breaches of digital security if the routers are not first wiped. Worse, the ACCC’s position is further confused by the fact that these routers can be flashed, making their destruction quite pointless.

Data discarded, but the routers still work

Since the June 30 cut-off, the routers are known to power up, and while bricked for internet access, can seemingly be reused with custom router firmware. The nature of the data collected by the routers for the MBA program was for measurement and customer registration, and this has apparently been deleted, according to emails sent to volunteers of the program.

While accurate figures are unknown, by December 2020 over 2,600 of the SamKnows SK-WB8 routers had been distributed by the ACCC, with 4,000 planned for release across the lifespan of the MBA program.

Attempts to contact the ACCC and Cisco to learn more about why volunteers of the MBA program are being encouraged to dispose of perfectly usable routers have been met with stock, non-committal responses.

The ACCC gave ArsTechnica a potted history of the device and the MBA program, as well as stating that “volunteers are encouraged to unplug their disabled whitebox and dispose of it in an environmentally responsible manner via free e-waste recycling services.”

Should you ditch your ACCC router?

If there is no reason to keep your ACCC-supplied router and have a replacement ready, it is important to factory reset the device before disposal. This ensures that any administrator passwords, ISP details, and custom network settings are deleted, keeping them from falling into the wrong hands.

Finding a safe disposal option will help to ensure the router is correctly dismantled. The ACCC has emailed recipients of the SamKnows SK-WB8 routers informing them of the correct disposal procedure, with a link to a live list of e-waste services.

However, if you have the time and inclination to flash OpenWRT, a guide explaining how to do this is available on the OpenWRT page for the SamKnows SK-WB8.

• ETH Zurich quantum chip sees superconducting qubit act as CPU and the vibrational modes of a fingernail-width acoustic resonator serve as quantum RAM
• The approach borrows from classical computer architecture as it completely flips the script on how modern quantum computing might store short-term data
• The team demonstrated a universal gate set and ran small instances of the quantum Fourier transform and period finding

A guitar string essentially stores a note based on how it vibrates, and if one plucks it differently, an entirely different note plays.

A team of researchers at ETH Zurich has leveraged the same principle to build a quantum chip that stores information by replacing the string with microscopic acoustic resonators.

This allows the chip to increase its working memory significantly, essentially increasing the storage capacity, a prohibitively expensive commodity in quantum computing, significantly.

A vibrations-based quantum storage play

ETH Zurich's research is led by quantum physicist Yiwen Chu, who used tiny mechanical vibrations to both store and process information. The vibrations, however, go far beyond the range of human hearing, happening inside a quantum chip where they essentially replace or complement the working memory of a quantum computer.

The study, published by the Hybrid Quantum Systems group, lists Professor Yiwen Chu, along with doctoral students Yu Yang and Igor Kladarić, as lead authors and focuses on replicating the division of labor seen in a classical computer.

A superconducting transmon qubit serves as the CPU, while the working memory (the quantum equivalent of RAM) is a high-overtone bulk acoustic wave resonator, or HBAR, whose many vibrational modes each serve as a memory slot.

The Qubit essentially swaps a quantum state from a vibrational mode (reads it, in classical computer terms), manipulates it (modifies it), and swaps it back (writes it). This makes for a unique configuration that most modern quantum computers do not follow, in which processing and storage are two distinct segments; most designs treat both memory and compute similarly.

The approach has advantages, however: acoustic waves have wavelengths roughly a hundred thousand times shorter than electromagnetic ones, allowing an entire quantum chip to be extremely small, as the research team states, even if the actual computer will be many orders of magnitude larger.

The chip has passed stress tests, including a proof of feasibility, which also included testing using two of the most commonly used methods to benchmark a quantum computer: the quantum Fourier transform and a period-finding algorithm.

The endgame here, as noted by the research team, is quantum random-access memory (QRAM), which would allow modern quantum computers to access a much larger store of quantum memory than current specifications allow. Whether this pans out depends on both the scalability of the approach and the computational power in play.

• Pentagon seeks an affordable drone capable of replacing many MQ-9 Reaper missions
• Reaper drones hit hard in Iran, and losses force demand for lower-cost unmanned combat aircraft
• New aircraft must combine long range with substantial payload capacity

The United States Air Force is examining a lower-cost unmanned aircraft concept after losing “dozens” of MQ-9 Reaper drones during the recent conflict involving Iran.

Those losses have intensified concerns about relying upon expensive aircraft in environments where increasingly affordable air defenses can destroy them.

With roughly 135 Reapers in service and each drone costing about $30 million, officials increasingly question whether existing loss rates remain sustainable.

Pentagon seeks a lower-cost drone with long range and heavy payload

Rather than pursuing a more advanced version of the MQ-9, defense planners are exploring a drone intended for larger-scale deployment.

The Defense Innovation Unit is seeking proposals for a Massed Modular Aircraft, or MMA, capable of performing many missions currently assigned to the Reaper.

According to the solicitation, the Pentagon believes dependence upon “exquisite” aircraft costing more than $30 million is becoming increasingly difficult to sustain.

The concept favours quantity alongside capability, allowing forces to continue operating even after suffering substantial battlefield attrition.

Unlike many smaller drones commonly associated with swarm operations, the proposed aircraft would retain significant reach and carrying capacity.

The solicitation calls for a payload of at least 2,800 pounds, compared with roughly 3,800 pounds carried by the MQ-9.

Requirements also include an unrefueled combat radius of at least 2,300 nautical miles and a one-way transfer distance exceeding 8,000 nautical miles.

The drone must travel at speeds above 200 miles per hour while remaining capable of operating from 6,000-foot runways and improvised airstrips.

Defense planners also want enough onboard power and cooling capacity to support diverse internal and external mission equipment.

The specifications mention 25kW of available electrical power and 5kW of cooling capacity for future mission systems.

Ambitious timeline aims for operational capability by 2031

The proposal places considerable emphasis upon autonomy, allowing a single operator to supervise several aircraft simultaneously during complex missions.

While no specific dimensions were included, the performance requirements indicate an aircraft broadly comparable in size to the MQ-9.

Officials have also not disclosed a preferred procurement price, though expectations suggest a figure substantially below the Reaper's estimated $30 million cost.

The timeline remains aggressive, with full-scale prototype flight testing expected within 21 months following contract award.

Initial Operating Capability is planned for fiscal year 2031, with 20 mission-ready aircraft delivered to an operational unit.

Recent combat experiences appear to have influenced the concept's development, particularly situations where defenders exhausted interceptors before attackers exhausted drones.

The solicitation argues that maintaining continuous MMA operations could pressure opponents into consuming costly defensive missiles at unsustainable rates.

“Keeping a constant airborne MMA presence to launch weapons, gather intelligence, perform electronic warfare missions, or relay communications will force an adversary to stay on the defensive,” the Defense Innovation Unit stated.

Via Defense News

The AI buildout is well and truly underway with OpenAI CEO Sam Altman leading the charge, having made various deals with companies including Oracle and Nvidia to guarantee the infrastructure needed to train future AI models is installed. But as this ensues, the spotlight has been thrown on how much energy these models will need.

Eating machines

Altman was speaking during an AI summit in India earlier this year when making the remarks to The Indian Express.

The OpenAI chief was being questioned about the substantial amount of energy that AI has already been consuming – and will be projected to continue consuming – for both training and inference. Bringing these models online, after all, and keeping them running require a huge amount of resources, not only in terms of the energy to power the data centers, but the water for cooling, and the components and resources in building the hardware.

Altman's defence hinged on the idea that people, too, require plenty of energy in order to reach utility — while AI can be trained much quicker. This view, in essence, frames machine intelligence as a like-for-like competitor with human intelligence.

Energy efficiency

Altman argued that humans are deeply inefficient, and that you should compare the total energy spent to create a human expert versus a machine expert. There's also an argument that energy efficiency of AI could improve over time.

But detractors were scathing in their criticism of this entire point of view – saying this framing doesn't take into account the fact that the human brain operates on roughly 20 watts of power. This isn't to mention the ethically gray and dehumanizing nature of the remarks.

• Environmental groups seek broader review before massive satellite constellations receive approval
• More than one million proposed satellites face increased regulatory scrutiny
• FCC is reconsidering satellite environmental review rules

Environmental groups have petitioned federal regulators to pause approval of orbital data center satellite constellations pending a full environmental review process.

Earthjustice recently filed a petition on behalf of DarkSky International, Environment America, and Public Employees for Environmental Responsibility, known as PEER.

Combined proposals from SpaceX, Starcloud, Blue Origin, and Cowboy Space could place well over a million satellites into low Earth orbit.

Why regulators are being asked to slow down

The petition asks the Federal Communications Commission (FCC) to prepare a Programmatic Environmental Impact Statement before approving any pending applications currently under review.

Such a review, required under the National Environmental Policy Act, would examine risks, alternatives, costs, and cumulative impacts together across every proposal.

Environmental groups argue the agency's current approach treats satellite licenses as automatically excluded from any detailed environmental scrutiny under existing federal rules.

They say that framework no longer fits proposals measured in hundreds of thousands, or potentially millions, of individual spacecraft rather than dozens.

The filing lists specific concerns, including rocket emissions, reentry pollutants, ozone depletion, orbital debris, and disruption to astronomy research conducted worldwide.

The petition specifically challenges the FCC's default assumption that these projects individually and cumulatively carry no environmental impact whatsoever on nearby ecosystems.

It further warns that light pollution and wildlife disruption cannot be properly assessed through isolated regulatory reviews conducted individually rather than collectively.

The petition states that these proposals compound risk "synergistically and cumulatively" in ways single-project reviews cannot capture on their own.

Industry ambitions collide with regulatory uncertainty

Backers of orbital computing describe their projects in sweeping, civilization-changing language while offering few environmental details in return for regulatory approval.

Companies including SpaceX, Blue Origin, Starcloud, and Cowboy Space have not publicly detailed environmental mitigation plans for their satellites currently under regulatory review.

MOL and Hitachi have separately explored floating data center concepts, showing wider commercial interest beyond traditional orbital satellite proposals currently facing regulatory review.

The FCC is separately reconsidering its environmental review rules, acknowledging rapid growth across the broader commercial space industry over the past decade.

If the commission agrees, orbital data center operators could face considerable regulatory delay before launching any additional hardware skyward into low Earth orbit.

Some industry analysts have already questioned whether orbital data center economics make sense given high launch and maintenance costs involved in space deployment.

Analysts note that environmental reviews of this scope could take years, delaying deployment timelines.

This could delay deployment schedules and extend regulatory timelines if a comprehensive environmental review becomes mandatory.

Whether the FCC ultimately requires a full review remains uncertain, given ongoing industry pressure and competing national security interests tied to space dominance.

Until regulators decide, the fate of orbital computing may depend as much on environmental politics as on rocket technology or launch capacity itself.

Via The Register

• AI hallucination can be weaponized, new report warns
• HalluSquatting is short for “adversarial hallucination squatting”
• GitHub Copilot, Gemini CLI, and OpenClaw are all affected

Your favorite AI service could be subverted to deploy code that turns your phone or PC into a botnet, according to researchers at Intuit, Technion, and Tel Aviv University.

The technique has been given the name HalluSquatting, a portmanteau of adversarial hallucination squatting, and is similar to typosquatting in that it relies on a mistake in order to distribute malicious code. While typosquatting might occur with the incorrect input of a website URL, HalluSquatting pivots on an LLM being unable to identify a resource or repository with 100% accuracy.

Relying on an LLM’s tendency to hallucinate repository resource identifiers, this weakness could be scaled up to conduct massive ransomware campaigns, botnets, and more.

Push-me-pull-you

Previous LLM-based malware operations have relied on pull-based attacks. In this scenario, a prompt designed to jailbreak or otherwise subvert the AI is (for example) placed on a website and the LLM encouraged to gather the information, thereby reducing its internal security.

What the researchers have shared in their paper, is that pull techniques are being combined with push attacks, which are traditionally executed as code injection.

The paper’s introduction summary states: “By preemptively registering hallucinated resources—a technique we call adversarial hallucination squatting (HalluSquatting)—we demonstrate remote tool execution and remote code execution at scale across a range of popular agentic LLM applications, which could be exploited to the establishment of a botnet.”

Once an attacker has identified the resource likely to be misnamed by an LLM, and squatted on it (to embed adversarial prompts), the work is done. All that remains is for a user to trigger the resource, the AI chatbot or agent to initiate the response, and the squatted resource will be accessed.

Promptware attack

Following this, the adversarial content held within the squatted resource is activated, triggering the tool invocation stage. This is the promptware attack, where attacker-controlled instructions are executed, with results potentially including turning the device you’re using into a botnet zombie.

LLMs such as the Cursor, Cursor CLI, Windsurf, GitHub Copilot, Cline coding assistants have been used in the testing of this avenue of attack along with Gemini CLI, and the OpenClaw, ZeroClaw, and NanoClaw AI assistants. The researchers successfully achieved remote tool execution (essentially remotely accessing and controlling the LLMs) and remote code execution (RCE, where malicious code is executed remotely).

Some mitigation is available, including LLM developers blocking fetch operations in favor of a search tool, and resource owners enforcing strict naming, perhaps in favor of globally unique resource names. However, these are will require collaboration by disparate parties, and may take a while to implement.

The risk of LLM-based malware is increasing, and some has already been spotted in the wild. Of these, the JADEPUFFER attack is perhaps the most notable, as it isn’t simply AI-based malware – it is a full ransomware attack run entirely by an LLM.

• The three man team is known as “Slopfix”
• It claims to be able to reduce AI generated codebases by up to 65% in size
• They aim to "refactor vibecoded codebases back to maintainability"

Vibecoding has a lot to answer for, not least some excessively large codebases. A new team of software engineers are collaborating to reduce the size of these cumbersome projects… with a $10,000 per week bill.

Slopfix is the name of the team (comprising a trio: Maciej, Kuba, and Krzysztof), but its aim is efficiency and functionality, rather than code golf, where code is reduced to the shortest possible length.

However, while this might seem like a noble task and a service worth paying for, Slopfix isn’t taking a stand against the use of AI. In fact, it is employing AI tools to detect the AI flab in your codebases.

Use an AI to catch an AI

Challenges around vibecoded projects have increased in recent months, as the limitations of the technology become apparent.

While using an AI to program based on your prompts and requirements is straightforward, agents habitually begin to lose context and logic once the project reaches a certain size or age. Once that happens, you’re looking at duplication, features breaking, and of course, the dreaded hallucination.

Slopfix is targeting companies that have adopted vibecoding, built huge codebases, and found that they’re running into issues. To find the problematic AI code, however, Slopfix is employing AI.

They state that a full “screen by screen, endpoint by endpoint” evaluation of the vibecoded app is made, which aims to find the duplicated functions, broken logic, and other issues. There’s also the promise of a two-week warranty for anything they break.

All of this is aided by Claude Code “on a very short leash” which Slopfix uses find problems. They clearly state that “the agent doesn’t get a vote.” Instead, they’re relying on their experience as developers to improve your code.

$10,000 seem a bit steep?

While the price might seem high, $10,000 for one successful week’s work for three seasoned developers shouldn’t really be a budget breaker.

The fee covers successful work only, and as the Slopfix website states, payment is in proportion to how much of the reduction target the team hits, with $10,000 being the price for hitting the target – it’s not the default fee.

However, there is a lot of preparation involved, and the analysis of your codebase is conducted free of charge. If they can't fix your project's issues, they'll let you know and refuse the contract.

As software consultancies go, Slopfix is an unusual case. But as the problems with vibecoded projects begin to become apparent, competing consultancies may begin offering similar services.

• Vyriy 15 FPV with The Fourth Law's TFL-1 AI guidance reportedly struck Russian logistics 68 miles (110 KM) away
• Ever-innovating Ukrainian drone industry continues to achieve economies of scale even as it becomes a growing threat to Russian advances
• With a payload capacity of 8kg and the ability to be equipped with a thermal imaging module as well as electronic warfare deterrence, it offers an interesting alternative to comparable fixed-wing drones that cost thousands of dollars

Basic FPV drones are hardly a new thing in a market flooded with hundreds, if not thousands of options that can cost as little as $100 to 200, but the Russia-Ukraine conflict might have upped the ante on affordability for a different kind of UAV that leverages the same tech: attack drones.

The Vyriy 15 is a self-styled "kamikaze drone" by the company that offers a stated strike range of 40-70km with up to a 8kg payload in tow which can be retrofitted with a thermal imaging module as well as an extended band VTX module to make jamming it harder.

With a control range of up to 30KM and a flight duration of 20 minutes (with a payload) and a cruising speed of 60-100 km/h, its not the most technologically advanced drone out there, but at its purported price tag of $500, it doesn't need to be.

An FPV strike record backed by AI

On the 10th of July, Yaroslav Azhnyuk, the CEO of Ukrainian autonomy developer The Fourth Law, announced on X what he called "a new FPV strike record": a Vyriy 15 quadcopter, flown by Ukraine's 5th Border Guard Detachment and fitted with his company's AI terminal-guidance module, had flown 110 km (68 miles) to strike a Russian logistics target.

This is both a significant achievement for Ukraine's domestic drone industry and a key indicator of how fast the Russia-Ukraine war has turned into one of attrition, with supply lines becoming increasingly targeted to prevent significant advances in either direction.

It also showcases how AI on the battlefield is shaping the conflict: the Vyriy 15 is, by default, a manually controlled drone that would otherwise need an operator or a relay to be closer to the theater of war.

The competition is American-made Hornets, fixed-wing drones that can cost upwards of $5,000, a 10-fold increase in cost for an already cash-strapped Ukrainian military that is increasingly looking towards localized solutions.

The optional AI module used to set the record is The Fourth Law's TFL-1, a machine-vision terminal-guidance module that operates on a fire-and-forget principle: once the operator visually designates a target, an onboard computer takes over the final approach, essentially countering Russian jammers that would otherwise disrupt a video link.

If Ukraine manages to mainstream such warfare in the future while cutting costs down to a tenth of what they do right now, reliably striking as deep as 100km into enemy territory while proving difficult to jam or costly to intercept, drones like the Vyriy 15 could signal an evolution in the modern battlefield even as aggression with low-cost drone swarms is already being rewarded in other conflicts such as the US-Iran war.

Geopolitical tensions are mounting, and nation states are employing new types of strategies to gain intelligence. A recent Five Eyes warning, for example, accused Chinese military intelligence officers of using professional networking sites and online job platforms to target individuals of interest.

In this specific case, the agents pose as recruiters advertising seemingly legitimate work to build relationships and, ultimately, get their hands on non-public information. Popular sites like LinkedIn, Indeed and Upwork have all seen this new type of attack take place.

At the same time, a parallel threat sees operatives applying for jobs within trusted organizations with access to intelligence, creating insider threats that experts warn AI might be mostly responsible for.

Generative AI, for example, can create documents, write applications and even supply live answers during real-time remote interviews, meaning that a small group of fake applicants can extend their reach much more quickly.

Rather than attacking existing workers, nation states are creating their own job candidates

Once inside an organization and with access to company tools like PCs, emails and other internal systems, nation state spies can then move laterally to acquire the information they sought.

Security experts at Exabeam warn that, because this technique is still evolving, it might not always be so easy to spot. Additionally, motives can differ, with Chinese intelligence operations typically seeking military, political or economic information. North Korean agents, on the other hand, tend to be tied to stealing money, which could also come with the side effect of data and intelligence theft.

Exabeam even observed this type of attack first-hand, when a North Korean-affiliated applicant used a false identity to apply for a job at the company. After passing technical tests, a video interview and other standard checks, the suspect’s laptop was quickly flagged for unusual activity.

In the following Q&A with AI Strategy and Security Research VP Steve Povolny, I discuss these new types of attacks, who’s responsible for stamping them out and what we can do to prevent similar incidents from happening more commonly.

• The Five Eyes alliance recently warned that foreign intelligence groups are using job platforms to recruit insiders. How significant is this threat, and what is driving its growth?

This is among the most serious access-driven threats facing cleared workers, and it keeps growing because the economics now favor the attacker.

Foreign intelligence services no longer need handlers and dead drops when they can post a job ad on LinkedIn or Upwork and let candidates self-select based on the access listed in their own resumes. Generative AI lets them run thousands of these conversations at once, drafting outreach and scoring which applicants sit closest to sensitive information without a trained officer.

The Five Eyes alert describes a scaled, automated funnel, and that scale is what makes it dangerous.

• A parallel risk runs alongside that warning: adversaries who secure employment directly rather than recruiting an existing employee. Which scenario presents the greater defensive challenge, and why?

The infiltration model gives defenders less to work with, which makes it the harder problem. When an adversary recruits someone already on staff, most of the suspicious behavior happens outside the company on platforms the employer never sees, yet the insider remains a known person with a verified identity and a real history.

When the adversary becomes the employee, the company has onboarded a fabricated person and handed them a laptop and standing network access on day one. No behavioral baseline exists, since everything that account does counts as a first. The deception also clears the controls most organizations trust, so the failure lands before any security tool gets a vote.

• Exabeam identified a North Korea-affiliated individual who gained employment at the company. How did the operative clear Exabeam's hiring process, and what first signaled that something was wrong?

He cleared it by performing well on the parts we test and forging the parts we verify. Applying under the alias Trevor Rothluebber, he aced the technical interview and take-home assessment, passed the video interview and cleared our standard pre-employment process including the background check and I-9 validation.

Our hiring team flagged a suspicion that he leaned on generative AI for live help during the video call, the first soft signal. The hard signal arrived the moment he logged into his corporate account. Our threat intelligence feed matched his username to activity previously associated with North Korean operatives and rated it high risk, and that single match reframed how the team read everything that followed.

Simultaneously, Exabeam’s platform detected a number of anomalies inconsistent with a brand new employee’s first day, and escalating in severity within hours. Incident response quietly isolated and reimaged his laptop before any real damage could be done.

• The candidate completed applications, interviews and assigned work without raising alarm. In retrospect, what indicators were present, and why did standard screening miss them?

The indicators existed, but they lived in places our screening was never built to read. The driver's license he submitted was either AI-generated or very badly manually modified, and the tell was physical. The image had unique aberrations, such as the ears in the photo which had an unnatural and pixelated modification an artifact that image generators still produce, and a reviewer skims past.

The live AI assistance during the interview was another, since his answers carried a fluency that did not match the natural hesitation you expect when someone reasons through an unfamiliar problem. Standard screening missed all of it because background checks and identity validation confirm whether documents are internally consistent and whether a record exists, and they never ask whether the human attached to those documents is real.

Further fabrication of documents such as I-9 were missed by a 3rd party identity verification company, and validation of (fake) job references was not properly identified.

• How did AI contribute to the deception? What did the fraudulent documentation involve, and what capabilities does AI introduce that traditional forgery methods lack?

AI showed up at nearly every stage. The fraudulent documentation centered on a forged driver's license we believe was generated rather than physically produced, paired with a stolen identity that gave the paperwork a real history to rest on.

During the interview the candidate appeared to have run an AI copilot feeding him answers in real time, and many of these tools now stay invisible to everyone else on the call even while the candidate shares a screen. What AI adds over traditional forgery is volume and believability together. A skilled forger could always produce one convincing passport, but the craft capped how many operations could run at once.

Generative tools remove that ceiling, so a single actor can fabricate convincing documents and coach themselves through a live technical interview across dozens of applications at once, and the forgery stopped being the bottleneck it used to be.

• The Five Eyes warning focused on China, while the Exabeam case involved North Korea. Do these actors share tactics and objectives, or do they represent distinct operational models that overlap on method?

They overlap heavily on method while running on different motives, which defenders should sit with. The Chinese operation the Five Eyes described aims at intelligence collection, pulling government and military insight out of people who already hold access.

The North Korean program that hit us and so many others in this industry is funded differently, since much of its purpose is revenue for a sanctioned regime, with intrusion and theft riding alongside the paycheck. The objectives diverge, yet the tradecraft has converged on one toolkit of fabricated identities, AI-assisted documents, manufactured professional histories and the patient relationship-building that lets an operative stay quiet.

When two adversaries with separate goals reach the same playbook, that tells you the playbook works and other actors are already watching.

• Conventional insider threat programs are built to detect employees who become compromised over time. How should organizations identify a candidate who was an adversary from the point of hire?

Our mindset must shift toward treating the moment of hire as the start of the highest-risk window rather than the end of vetting. Traditional insider programs watch for drift, the employee who gradually turns after a financial shock or a grievance, so they depend on a baseline built over months.

A candidate who was hostile from day one never produces that baseline, which forces you to scrutinize the earliest behavior most closely. In our case, the catch came from putting new accounts under enhanced monitoring and letting an AI agent correlate scattered signals that no single alert would have justified escalating.

The working principle is to give hiring workflows and new-hire activity the same suspicion you already apply to production access.

• Where should accountability for this threat reside within an organization? Is it a security function, an HR function, or a gap that persists because ownership is unclear?

Accountability most often lives in the gap right now, and that gap is exactly why the threat works. Hiring sits with HR and talent acquisition, who are measured on filling roles quickly and are not equipped to run identity verification at an intelligence-grade level.

Detection sits with security, which usually gains no visibility into a candidate until that person already holds a badge and a laptop, and the adversary exploits the seam between the two.

The workable answer is shared ownership with a clean handoff, where security sets the identity and behavioral standards hiring must meet and stays involved through the first weeks of employment rather than inheriting the problem once onboarding closes.

• Many mid-sized companies lack dedicated threat intelligence resources. What practical measures can such organizations implement to reduce their exposure?

Useful defense does not require a dedicated threat intelligence team. The interview itself is the cheapest control available, and small changes make it far more revealing.

Underspecifying a problem on purpose shows whether a candidate asks clarifying questions like a real engineer or simply produces a confident answer and switching the problem partway through tests whether they adapt or whether something is feeding them responses.

Asking for an external webcam that shows the workspace instead of a shared screen removes one of the easiest hiding spots for an interview copilot. Beyond hiring, the highest-leverage move is placing every new employee on a watchlist for closer monitoring through their first weeks, which costs configuration time rather than budget.

Even a basic, low-cost threat intelligence feed would have surfaced the username match that broke our case open.

• What is the most contested prediction on this issue, one that many security leaders would currently dispute?

My contested prediction is that within a couple of years the verified human interview, run live and in person for any role with meaningful access, returns as a security requirement. Many security leaders will fight that because it breaks the remote-first hiring model they spent years optimizing.

The objection I expect is that it does not scale and shrinks the talent pool, and those concerns are legitimate. My counter is that the economics have already flipped for high-access roles, since the cost of onboarding a single fabricated adversary now dwarfs the friction of one in-person verification step.

The deeper claim underneath it is that remote identity verification as we practice it today is no longer reliable for sensitive positions, and AI is what made it unreliable. Most security leaders are not ready to say that out loud yet.

• Group‑IB warns of scams exploiting Celine Dion’s concert comeback, with fraudsters selling duplicate Ticketmaster tickets and spoofing sites like AXS and Paris La Défense Arena
• Scammers embed themselves in Facebook fan groups and marketplaces, even using voice messages to build trust and make fake offers seem legitimate
• Fans are advised to only buy from official distributors, verify tickets in person if using resellers, and contact banks to dispute charges if scammed

Celine Dion is back, and hackers are already trying to exploit the fact for their own financial gain, experts have warned.

A report from security researchers Group-IB has claimed there are numerous scam campaigns all across the internet and social media, looking to exploit gullible fans and steal their money.

Its aptly named “The Scam Will Go On” report said it saw scammers lurking in Facebook Groups, Facebook Marketplace, and other fan-centric spaces, offering concert tickets for sale. The tickets themselves, hosted on Ticketmaster, are valid. However, the scammers only have a few tickets which can be redeemed by the first person who reaches the venue. Everyone else will be denied entry, since their tickets will already have been used.

How to avoid getting scammed

But that’s not the only scam. Some people don’t want to pay an unknown third person via wire, and would prefer to purchase the tickets directly from a service.

For those people, the scammers created entire websites, spoofing ticketing distributors such as AXS and Ticketmaster. Group-IB also saw fake websites spoofing Celine Dion and Paris La Défense Arena, the stadium where the concert will take place.

“We see that such an event generates excitement and provides scammers with another opportunity to make a fortune at the expense of unsuspecting fans,” Group-IB warned.

“Scammers are using increasingly sophisticated techniques, such as embedding themselves into social networking fan groups and speaking directly to their victims via voice messages to make the interaction more personal and gain their victims’ trust more easily. Furthermore, official ticketing platforms are being misused to make scams seem legitimate.”

The researchers recommend fans only visit official websites and those of official distributors, and if they absolutely must buy from a reseller, to make sure they’re purchasing a physical ticket, in person. Those that fell for the scam should call their bank and lodge an objection on their credit card.

If you're looking for a reliable and cheap mini PC, the GMKtec G10 is our top choice. For general day-to-day work, with the scope to upgrade, there's plenty to like about the machine we called "a terrific little system" for the price.

Right now, the GMKtek G10 mini PC is $288 (was $400) at Amazon — a $112 saving on a genuinely capable budget mini PC. Meanwhile, in the UK, the G10 is now £260 (was £319) at Amazon.

Today's top mini PC deal

In our review, we found it well-balanced for light office tasks, web browsing, and streaming. If that's the goal for your setup, it's ideal. For more powerful machines designed for demanding graphical workloads, we've selected the top mini PC deals for video editing and gaming.

So, what's on offer here? The Ryzen 5 3500U isn't new silicon — it's a 2019-era Zen+ chip on a 12nm process that still holds up very well against the sluggish Intel N-series chips (N100, N150) that populate most budget mini PCs at this price. The four cores and eight threads, courtesy of AMD's multithreading implementation, give it a genuine edge in multitasking and anything that touches video decoding or light image processing.

Where the G10 stands out is memory and storage. 16GB of dual-channel DDR4 and a 512GB NVMe SSD is a generous starting configuration at this price, and GMKtec gives you real room to grow: two SO-DIMM slots support up to 64GB of RAM, and two M.2 2280 slots support up to 16TB of combined storage. Having two drive bays also makes cloning an existing installation onto a bigger drive simple if you outgrow the included SSD.

2.5GbE Ethernet is the other feature worth calling out, since most rivals in this price bracket still ship with a single Gigabit port. If you're setting this up as a home server, NAS front-end, or Proxmox node, that faster wired connection is a genuine practical upgrade over the competition, and the Ryzen 5 3500U's AMD-Vi support means virtualization software runs without issue.

The Radeon Vega 8 graphics are dated by 2026 standards and won't handle demanding modern games, but they're meaningfully more capable than Intel's UHD graphics for everyday tasks like 4K video playback, light photo editing, or older and indie titles at lower settings.

Good to know: this machine needs a bit of patience out of the box. We found the G10 performs noticeably better after a trip into the BIOS to switch from Balanced to Performance mode, and the fan does get audible under sustained load in that mode. It's not a plug-and-play polished experience — with around 30 minutes of setup, you'll get this chip at its best.

For a cheap mini PC with genuinely generous memory, storage headroom, and fast networking, the G10 is a solid budget pick that features in our guide to the best mini PCs.

If you want home backup power without a bulky, heavy unit taking up a corner of the garage, this is a genuinely good price for it. Right now, the Anker Solix S2000 is $600 (was $1199) at Amazon — a massive 50% off on a 2kWh power station built specifically for keeping essentials running through an outage.

The headline feature here is the size-to-capacity ratio. At 35.7lbs and roughly the footprint of a large kitchen bin, the S2000 packs in 2,010Wh of LiFePO4 capacity in a chassis around 30% smaller and 25% lighter than typical 2kWh power stations. That matters if you're planning to keep this somewhere specific in the house rather than store it in a garage.

Today's top portable power station deal

Anker's OptiSave system is the other notable piece of engineering. It drops the unit's active idle power draw down to around 6W with the AC outlets enabled, and roughly 2W with everything switched off — well below the 14-20W idle draw typical of competing 2kWh stations. Since a backup power station mostly sits idle waiting for an outage, or cycles a fridge compressor on and off, that lower standby drain translates into meaningfully more real-world runtime rather than just a better spec sheet number.

Port placement is a small but genuinely useful design choice: two AC outlets sit on the back for always-on gear like a fridge or router, while three more AC outlets plus USB-C and USB-A ports sit on the front for everyday devices. That split keeps cables from getting tangled if you're running multiple things at once, and the 10ms UPS switchover means a connected computer or router won't even blink when the grid drops.

The 10,000-cycle LiFePO4 battery is rated for roughly 15 years of typical use, which is close to double the cycle rating most mainstream power stations in this capacity class advertise — a genuine long-term value argument if you're planning to keep this around rather than replace it after a few years.

Important notes: the S2000 caps out at 1,500W continuous output, so it's not the right choice if you need to run power tools, space heaters, or other high-draw appliances — Anker's own larger Solix units handle that heavier lifting. Solar input is also capped at 400W, lower than some rivals that accept 600-1,000W, so it's better suited to topping up with a single panel than building out a larger solar array. And with only two USB-C ports and one USB-A port, it's clearly built around AC backup duty first, device charging second.

For anyone who wants dependable home backup without the bulk of a typical 2kWh unit, the SOLIX S2000 at $599.99 is a strong buy.

Also consider: More portable power station deals

If you're anything like me, one of the biggest frustrations with printing is the cost of ink. Over time, it can easily cost more than the printer itself, making refillable ink tank models an increasingly popular choice. They're built to keep running costs down while reducing the hassle of replacing ink cartridges on a regular basis.

So this deal immediately stood out to me. The HP Smart Tank 5101 all-in-one printer is down to $190 (was $260) at Amazon, saving you $70.10. In the UK, it's known as the HP Smart Tank 5107, and it's discounted to £150 (was £210) at Amazon.

Instead of relying on traditional cartridges, this model uses ink tanks that hold enough ink to print thousands of pages before needing a refill. Best of all, it comes with up to two years of ink in the box, so it will be ages before you need to spend out.

Today's top HP Smart Tank printer deal

Refilling the ink is easy because the bottles slot directly into the tanks and empty automatically without needing a squeeze or creating a mess, removing one of the biggest frustrations people have with refillable printers.

Printing speeds reach up to 12ppm in black and 5ppm in color, while the built-in scanner and copier mean it can handle far more than simple printing tasks. Wireless connectivity over 2.4GHz Wi-Fi lets everyone in the house or small office send jobs without plugging into a computer.

HP also says the Smart Tank 5101 continues producing sharp text and rich colors without the complicated maintenance routines often associated with ink tank printers.

Auto Power On helps reduce waiting around by waking the printer the moment a job is received, while borderless printing adds extra flexibility for photos, invitations, and creative projects.

Another useful addition is HP AI formatting, which automatically cleans up web pages and emails before printing, removing unnecessary content and awkward page breaks, stopping you wasting paper on unwanted content.

For the price, the Smart Tank 5101 is not only much cheaper than normal, it also comes with enough ink to keep many households printing for a very long time.

If your current printer burns through ink far too quickly, or you're thinking of buying your first all-in-one model, this is a terrific choice to get you started.

For more options, check out our round up of the best ink tank printers you can buy, as well as the best home printers.

Also consider

• Shop ink tank printers at Amazon

• Squarerspace launches new limited-release tool
• Features include a limited-time cart, purchase quantity limits, and low inventory badges
• Available globally now, although limited to some subscription plans

From game consoles to toilet roll, there are countless examples of how scarcity can make a product more desirable. Whether a smart marketing tactic or a simple issue with supply, limited availability drives FOMO. Now, Squarespace, one of the biggest website builders on the market, has launched new tools to help users with limited releases and product "drops."

So, whether you want to drive anticipation behind a one-off t-shirt design or want to cap spaces in an exclusive yoga class, Squarespace aims to equip sellers with everything they need to turn a limited availability product or service into the hottest buy on the internet.

Driving urgency

Key features of Squarespace’s new tool include:

• A reserved cart feature that holds items during checkout for a limited time, with a countdown timer, helping drive urgency and replenishing stock if the user fails to check out
• Inventory management tools that allow merchants to set quantity limits, preventing individuals from sweeping inventory, whilst also driving a sense of exclusivity
• Seamless checkout, offering a shortcut to purchase, so customers don’t have to navigate away from the store page to view their cart and check out
• Hype generation tools, including low inventory badges and email campaigns, to help merchants build anticipation and signal scarcity.

"We're seeing entrepreneurs shifting away from always-on discount models and toward highly curated, exclusive releases that celebrate the value of their work," said Kevin Doerr, President of Squarespace. "With our new limited release selling tools, we are putting entrepreneurs in control of when and how they sell – reducing the operational burden and giving them everything they need to build hype, handle surges in demand, and deliver a seamless checkout experience.

According to a press release shared with TechRadar earlier this week, limited release tools are available globally, although availability does vary depending on your subscription plan.

• Attackers are spoofing LastPass and Bitwarden with phishing emails from fake newsletter domains, tricking users into signing bogus DocuSign documents
• Victims are redirected to malicious “compliance” domains flagged by Microsoft Defender and Cloudflare, already taken offline
• Neither password manager was breached; this is domain spoofing, and users are urged to verify sender addresses and domains before clicking links

Criminals have been found impersonating popular password managers LastPass and Bitwarden online in an attempt to trick users into sharing their login credentials, and thus access to a treasure trove of passwords and other secrets.

LastPass recently issued a warning to its customers, raising awareness of the ongoing phishing campaign.

However the scam also now seems to have spread to other password managers, with Bitwarden customers also apparently being targeted.

Passwords are safe

In the campaign, LastPass users received emails from the address “hello@lastpassnewsletter.com”.

This address does not belong to LastPass, and is in no way affiliated with the password manager. In the message, the victims are told that the company’s security policies have been updated, and that they should navigate to a specific landing page and sign a DocuSign document.

The email comes with a ‘Review & Access Terms’ button which, if clicked, redirects the victims to lastpasscompliance[dot]com, yet another domain unaffiliated with the password management platform.

BleepingComputer claims this domain has already been flagged as malicious by both Microsoft Defender for Office 365, and Cloudflare and is currently offline.

Digging deeper, the journalists uncovered another campaign, almost identical, but now targeting Bitwarden users. In this case, the victims were being mailed from the “hello@bitwardennewsletter.com” addresses and were being redirected to bitwardencompliance[dot]com. Identical methodology, just slightly personalized.

It is important to note that neither LastPass nor Bitwarden were compromised as part of this attack.

The companies’ infrastructure is intact, and the passwords are safe. This is a typical domain spoofing attack in which the crooks purchase a domain similar to the legitimate one, in hopes that the victims won’t spot the difference.

As usual, the best course of action is to always be skeptical of incoming emails, and to double-check the domains and email addresses from which they are sent. It is also good to cross-reference these emails with any older messages that are proven to be authentic, to see if the domains and addresses match.

Mechanical hard drives (HDDs) still make the most sense when you need huge amounts of dependable storage, especially in a NAS that runs around the clock. If you're looking to expand your setup without overspending, I've found a terrific deal for you at Amazon.

The 8TB WD Red Plus NAS is currently down to $259 (was $355) at Amazon - a solid $96 off the usual asking price. Considering the 4TB version we tested is a little under $200 right now, that's a very welcome discount.

That's a great price, then, for a drive built specifically for always-on network storage, making it a great opportunity to add capacity to an existing NAS or form the basis of a brand new one.

Today's best WD NAS deal

Reviewing the 4TB WD Red Plus, our drive expert Mark even noted that the 8TB and 10TB models "offer the best value."

8TB of storage is plenty of room for family photos, videos, backups, documents, and media libraries.

Unlike desktop hard drives, this model is built for RAID-optimized NAS systems and uses NASware firmware to improve compatibility with supported enclosures.

It's intended for small and medium business environments, although it's just as useful for home users.

The drive spins at 5640RPM and combines SATA 6Gb/s connectivity with a generous 256MB cache to deliver dependable performance for file transfers, backups, and media streaming.

It also uses CMR recording technology, which many NAS owners prefer because it delivers more consistent write performance than SMR drives during heavier workloads.

Western Digital rates the drive for workloads of up to 180TB per year, making it suitable for systems that see regular activity not just occasional backups.

Support for NAS systems with up to eight drive bays also gives you flexibility if your storage needs is likely to grow over time.

Although SSDs dominate for operating systems and applications, traditional hard drives offer far better capacity for the money and saving almost $96 makes this one of the better prices I've seen recently.

For more picks, take a look at guides to the best NAS devices, best NAS hard drives (including the WD Red Pro) and best NAS and media server distros.

Whether you're an office worker or a student, I always find now is a great time to pick up a laptop for work and study that's actually worth the money.

MacBooks might be expensive, Chromebooks can be underpowered, but there are plenty of big-brand Windows 11 machines under $600 right now that can handle productivity tasks, schoolwork, creative projects, and everyday use without issue, and will stay useful well beyond the next school year.

These are the three deals that stood out to me right now.

Today's best laptop deals

My top pick overall here is HP's OmniBook 3, which is currently $599.99 (was $949.99) at Best Buy. It runs on Qualcomm's Snapdragon X processor with up to 32.25 hours of battery life, making it an excellent choice for office professionals and students planning to spend many hours away from a power outlet.

It comes with 16GB of memory alongside a 512GB SSD, making it far better equipped for multitasking and running Microsoft's growing collection of AI features. The laptop also features a 14-inch 1920 x 1200 touchscreen, Wi-Fi 6E, and HDMI 2.1 for hooking up an external monitor.

Another great choice is Asus' Vivobook 14. Also on sale at Best Buy, the model is priced at an even more affordable $548.

It combines the same Snapdragon X processor as the OmniBook 3 with 16GB of memory, a 512GB PCIe 4.0 SSD, and a 14-inch 1920 x 1200 IPS display.

Although it skips the touchscreen, it does keep the Copilot+ PC AI features, 45 TOPS NPU, WiFi 6E, HDMI 2.1, USB4, and impressive 29-hour battery life.

At 3.28lb, it's also the lightest laptop here, making it a great fit for students carrying it between classes. If you don't need touch support, I'd say this is the best all-round buy of the three.

Finally, there's the Dell 15 Laptop for $550 (was $750) at Dell. Powered by a 13th Generation Intel Core i7-1355U processor, it comes with a 512GB PCIe NVMe SSD, and 15.6-inch Full HD 120Hz IPS display which will make it useful for recording notes, writing essays, multitasking, and so on.

The laptop includes Wi-Fi 6, USB-A, USB-C and HDMI ports, as well as an SD card reader, and a headphone jack. My biggest reservation is that it only comes with 8GB of DDR5 memory, which really isn't enough in 2026. You can upgrade the laptop to 16GB of RAM if you want, but that will cause the price to skyrocket to $949.99.

For more top choices, we've also tested the best student laptops, best business laptops and the best laptops for engineering students.

• ArcticWolf uncovered 292 malicious GitHub repositories spoofing legitimate tools and products, delivering a new BoryptGrab infostealer variant
• Malware steals from 19 browsers, 32 crypto wallets, messaging apps, Steam, and Windows Credential Manager, and uniquely bypasses Chrome’s App‑Bound Encryption via code injection
• Most repos have been removed, but some remain active; GitHub’s popularity makes it a prime target, underscoring the need to vet code before use

Russian actors have reportedly created hundreds of malicious GitHub repositories masquerading as legitimate software but acting as a dangerous infostealer.

Cybersecurity researchers ArcticWolf discovered the campaign after finding their own products spoofed as part of the attack.

In total, the researchers found 292 fake repositories, spoofing things like security products, developer tools, macOS utilities, games, and more. Each repository contained a README file with the download URL.

Obviously malicious

Victims who download the program get a variant of the BoryptGrab infostealer family that grabs data from 19 browsers (passwords, cookies, payment information), 32 cryptocurrency wallets, Telegram, Discord, and Steam sessions, credentials for Meta’s Max, data from Windows Credential Manager, and more. It can also exfiltrate files from Desktop and Documents, and grab screenshots.

While most of the features can be found in other BoryptGrab variants, this one is unique in a sense that it can bypass Chrome’s App-Bound Encryption through direct code injection into the browser process.

While it hasn’t been specifically said that the threat actors are Russian, the compressed data is later sent to a Russia-based command-and-control (C2) infrastructure.

What’s also worth mentioning is that the malware is not designed to last. It has no anti-analysis layer, and doesn’t even try to hide itself in any specific manner. It does not establish persistence and simply tries to grab as much sensitive data as it can on the first attempt.

The attack, which seems to have started in the final days of June, is almost thwarted now, since most of the malicious repositories have been removed from GitHub. Citing “researchers”, BleepingComputer reported that several dozen still remain active, though.

Because of its importance and popularity in the open-source community, GitHub is currently one of the most targeted platforms on the internet, which is why it’s important to double-check and vet every piece of code before it’s applied to a project.

• AI PCs are emerging as a viable option to run local AI without unpredictable costs
• One-time PC cost alleviates the need to fork out for cloud token fees
• Broader research reaffirms rising popularity of smaller models

New Gartner data has claimed now could actually be a good time to buy AI PCs, as cloud computing faces numerous challenges in a rapidly-changing business world.

Data center construction is slipping behind demand as supple chains strain and local communities oppose new projects, meaning that metered compute could end up costing some companies more than they'd bargained for.

By shifting some of their AI processing locally, companies could be able to avoid some of those extra monthly costs with a one-time purchase of a more powerful PC as part of their regular refresh cycles.

AI PCs present an ideal hybrid compute model

While AI PC adoption started pretty slow with companies struggling to understand the benefits, they're now being seen as a cloud fallback option rather than a primary benefit in their own right.

With AI usage increasingly sharply and unpredictable token consumption hitting companies hard, forecasting monthly costs is a new major challenge that many face.

Small language and reasoning models, including specially trained models for individual business use cases, ultimately need fewer resources than leading frontier models, allowing them to be run locally as part of a broader hybrid approach.

Gartner predicts that speech and chat, text generation, image and audio generation and more could soon shift to workers' PCs, with only the most intensive tasks routed via hyperscaler data centers.

By as soon as 2029, the company's researchers anticipate that around one-third (30%) of enterprises could use AI PCs to reduce cloud AI token costs. By 2030, 70% of corporate PCs could be able to run some GenAI tasks locally.

Omdia researchers also noticed a shift in AI model usage, with smaller and medium models proving popular, with domain-specific tasks not needing the full breadth of compute.

"Older GPUs are retaining value and remaining in service, as they continue to offer a cost-effective option for small and midsized model inference and disaggregation," Senior Principal Analyst for Advanced Computing Alexander Harrowell said.

Via The Register

As the common phrase goes – ‘patience is a virtue’. The ability to endure, wait and yield to time without complaint is widely seen as a positive personality trait. Yet patience is situational, and everyone has a limit.

In a customer service context, people will wait longer when the stakes are high, or when they trust that delays will lead to better security, accuracy or care. If they’re dealing with another human, they also tend to be far more forgiving.

Alongside that limited patience, customers are also increasingly distracted. The average Brit now sits on over 1,000 unread emails in their inbox, according to recent research, alongside juggling around 25 non-work notifications a day.

People are not short of communication, but saturated. Attention is at a premium, and anything irrelevant is quickly deprioritized or disappears entirely.

This is particularly acute for the 36-55-year-old cohort, which boast the highest unread email counts, and the greatest pressure to stay online for work. Two-fifths of 36-50-year-olds say they feel more disconnected than ever, despite increased digital interactions.

Balancing careers, caregiving, mortgages and performance metrics, the attention and mental load on this group is significant. If brands are adding irrelevant noise and difficulty into these environments, tolerance is low.

This simultaneous feeling of burnout and impatience is heightened by the fact that the instant nature of the attention economy, and the incentives of digital platforms, has rewired expectations.

If there is no sense of progress or urgency conveyed with a digital service, then reassurance falls away and the chances someone will give up on the process rapidly increases. Clear, proactive explanations, combined with customer experience design that maintains continuity across every channel and interaction are vital.

Setting clear expectations

Digital saturation and the attention economy have also meant that first impressions are decisive. If in those initial moments of engagement there is any sense of delay or stalling, people will reconsider if they need the product or service in the first place.

These early moments are therefore a critical opportunity to build seamless customer journeys, closing these ‘patience deficits’ by finding and fixing areas that cause frustration. When there are unavoidable moments where waiting is required, brands must turn these into opportunities to strengthen loyalty and trust.

Customers want assurance that every click, confirmation or verification step serves a purpose. Uncertainty is the true enemy of patience, so give customers better visibility and control to help build trust.

Other industry research found that 25% of consumers there think ‘transparency about what’s happening and why’ is one of the most important things to them in a digital customer experience, with 45% valuing clear instructions and easy-to-follow steps. If there’s an action they don’t need to take or unnecessary duplication, remove it.

Poor design can lead to digital impatience, and exacerbate the overload of digital admin. Ensure that digital and automated channels are fast, reliable and transparent when it comes to waiting times. Competitive advantage will come from being the most considered, not the loudest.

Turning impatience into opportunity

At the same time, when customers know brands are protecting their data, there is an opportunity to earn patience through careful application of friction. Digital speed bumps like two-factor authentication, framed in the right way, can become symbols of care, not inconvenience – so in designing the customer journey, brands should take advantage.

The reality of digital patience is made more complex by the fact that more people are interacting with machines for customer service than ever before. AI agents, designed and used effectively, have raised expectations – but when AI gets stuck, misunderstands intent or is otherwise poorly designed, it backfires and becomes counterproductive. In those moments, many of us prefer to just talk to a human.

The purpose and context of interactions matter, so organizations must take care to match automation to the complexity and stakes of each task. Examine which tasks are best for AI agents to take on, and be clear about when AI is in use, explaining its role in plain language.

For those tasks which require more reassurance, empathy and accountability, human alternatives may be preferable. In these situations, we tend to be slightly more patient – 84% will stay patient on the phone with a real person, perhaps because we know we’re more likely to be able to connect on a human level, and get some understanding and reassurance in response to our enquiry.

If a service is too slow, however, then more than half of consumers say that delays lead them to think less of a brand, leave negative reviews, or warn others to steer clear.

Brands should provide the choice of human support and digital self-service tools in a way that reflects these dynamics. When switching between the two, brands should also carry context over, so that customers never need to repeat themselves.

Digital patience is a precious resource. Poor design which frustrates and builds impatience can see it easily lost, but with care, it is equally easy to earn. When brands anticipate frustration and design for reassurance, they convert fleeting attention into enduring trust, and find opportunities to strengthen relationships and loyalty.

We've featured the best ecommerce platform.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Every day, millions of people type a web address into their browser, usually in a flurry of rapid keystrokes, and arrive exactly where they intended.

However, a small but significant number of people don’t.

They might miss a letter, type an extra letter in their haste, or get some letters mixed up. Instead of hitting linkedin[.]com, they hit linkdein[.]com. Those mistakes gave way to one of the internet’s least glamorous destinations – the parked domain.

Most internet users have encountered them at some point, even if they didn't know what they were looking at. Typically, a parked domain would just be a sparse, messy page filled with adverts and a search bar with very little else.

They existed because someone, somewhere, recognized that in the early days of the internet a percentage of users would inevitably mistype a popular website – so they registered the most similar-looking domain names for themselves in a move known as “typosquatting” and earned advertising revenue from the resulting traffic.

If just 0.1% of the millions of people accessing amazon[.]com accidentally went to the amazn[.]com domain they’d bought, that’s still a worthwhile payday. It was a mundane corner of the digital economy, built on convenience, coincidence, and the occasional typo.

History can be a harsh teacher, but it can also sow complacency. In 2026, a lot of security teams still regard parked domains as little more than lazy digital billboards – inconvenient and annoying, but not a meaningful security concern.

However, the infrastructure surrounding parked domains has evolved considerably from the amateurish, pop-up-ridden advertising pages of the early internet. What was once a simple case of opportunistic domain monetization now sits inside a far more complex ecosystem of advertisers, brokers, and traffic distribution networks.

In many cases, a user's accidental visit no longer ends on a parked page at all. Instead, it triggers a journey through a chain of intermediaries operating largely out of sight. Somewhere along that journey, legitimate advertising can give way to fraud, scams, and malware distribution.

In other words, one of the web's most familiar and overlooked mechanisms has become one of the most lucrative and insidious vehicles for cybercrime.

From Mistype to Malware

The transformation of parked domains from digital curiosities into cybersecurity risks has been subtle, and that’s one of the reasons it’s so dangerous. For decades, the model followed the same patterns – a user would land on a parked domain, see a collection of banners, click on something accidental or otherwise, and generate a small amount of revenue for the domain owner.

It was cynical, but at least it was transparent because users could at least see where they had ended up and decide for themselves what to do next – usually just close the tab and go where they meant to. The only real danger here came from the occasional misleading or malicious ad rather than the mechanics of the domain itself.

Today things are different. Changes within the online advertising industry, including tighter policies around traditional domain monetization, have encouraged cybercriminals and fraudsters to try new approaches to keep the train of monetization moving.

Increasingly, users who arrive at a parked domain don't encounter a parked page at all. Instead, they’re immediately redirected elsewhere through a process known as “zero-click advertising”, sometimes referred to as direct search.

What appears to be a simple typo can trigger a rapid auction in which a user's visit is bought, sold, and passed between multiple advertising partners before they ever see a destination website. Most of this activity unfolds in fractions of a second and entirely beyond the user's view, and while many of those transactions remain legitimate, the sheer complexity of the ecosystem creates opportunities for abuse.

Somewhere within that chain, traffic can be acquired by actors whose interests extend far beyond advertising revenue, opening the door to scams, malware, fraudulent software, and a host of other malicious outcomes.

The Malvertising Economy

One of the reasons parked domain abuse is still underestimated and difficult to pin down is that the attack path rarely follows a straight line. When most people imagine a cyberattack, they picture a malicious website waiting at the end of a link, ready to ensnare an unsuspecting user.

But in this case, by the time a user reaches the content they're ultimately shown, their traffic may have already passed through a maze of advertising exchanges, brokers, redirectors, and cloaking services.

Each participant sees only a fragment of the overall journey, making it remarkably difficult for the “good guys” to determine which “bad guys” are responsible for what. It’s a little like trying to investigate a crime scene where the evidence constantly rearranges itself.

The cowardly threat actors involved in this type of cybercrime exploit this ambiguity. They use sophisticated cloaking techniques which allow them to examine visitors before deciding which content to serve up – where are they based? What kind of browser are they using? What operating system is their device running?

A security researcher in California might see a harmless landing page, while a finance broker in London might be served up a credential harvesting scam. This selective delivery makes malicious activity harder to detect and even harder to reproduce.

What’s worse, parked domain abuse is rarely aimed at a specific industry or organization. The actors deploying parked domains are usually financially motivated, and their primary interest is in acquiring traffic, so they’re not going to discriminate.

Once they’ve ensnared a victim, they become a commodity moving through an invisible marketplace where every click has value and every redirection creates another opportunity for exploitation.

The blind spot in traditional security

So where does all of this leave defenders? Parked domain abuse doesn’t behave like a conventional cyber threat. While security teams are used to investigating suspicious websites, malicious files, or compromised accounts that leave a relatively obvious trail, parked domain campaigns are different because the underlying traffic distribution is constantly changing.

The same typo domain can send one user down an entirely different path than the next. By the time an analyst attempts to recreate what a victim experienced, the route may no longer exist and any “evidence” has effectively evaporated. How do they defend against something they can't see or recreate?

One thing is guaranteed – regardless of how many redirects, intermediaries, cloaking systems, or advertising platforms sit between the initial typo and the final destination, every step in the journey depends on the domain name system (DNS). Often described as the internet's address book, DNS is responsible for translating domain names into the destinations users ultimately reach.

Put simply, each lookup leaves behind a breadcrumb that helps reveal relationships that would otherwise remain hidden, and that visibility has allowed researchers investigating typosquatted versions of well-known domains to follow the trail beyond the initial deception. Patterns start to emerge between seemingly unrelated cases of malware, involving the same parking providers, cloaking services, and traffic distribution infrastructure.

By examining historical DNS records and mapping the relationships between domains over time, it has become possible to connect incidents that appear to be unrelated and expose the networks operating behind them. Instead of playing “whack a mole” and chasing surface level domains, DNS mapping has allowed defenders to target the entire machine.

The greatest danger posed by parked domains isn't the typo itself, but the assumption that the infrastructure behind that typo is benign. For years, parked domains occupied a strange corner of the internet, largely ignored by security teams and rarely considered worthy of serious scrutiny.

But today, they offer cybercriminals something far more valuable than advertising revenue – access to legitimate systems, trusted business models, and vast streams of user traffic that can be manipulated and monetized at scale.

As threat actors continue to refine their use of cloaking, traffic distribution, and advertising networks, the distinction between legitimate online activity and malicious activity will become increasingly difficult to spot from the outside.

Protect yourself against malware with the best antivirus software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• ESET discovers 11 vulnerable UEFI shim bootloaders signed by Microsoft, allowing attackers to bypass Secure Boot and deploy malicious bootkits
• Any UEFI system trusting Microsoft’s 2011 third‑party certificate could be exposed, potentially billions of devices; attackers can bring old trusted shims to new systems
• Microsoft has revoked the vulnerable shims, and users should apply the latest UEFI revocations (Windows auto‑updates, Linux via LVFS) to block exploitation

Cybersecurity experts from ESET have discovered 11 vulnerable UEFI shim bootloaders, all signed by Microsoft, which could allow threat actors to exploit ancient vulnerabilities and bypass UEFI Secure Boot, deploying all sorts of malicious bootkits.

A shim is a small, intermediary bootloader that works as a bridge between a computer's firmware (UEFI) and the operating system's bootloader. Its primary purpose is to allow operating systems to work with UEFI Secure Boot without having Microsoft sign every Linux bootloader individually.

Any UEFI-based machine that trusts the Microsoft Corporation UEFI CA 2011 third-party UEFI certificate authority (CE) certificate, regardless of the operating system, was said to be vulnerable to the shims (versions 0.9 and older). That would put the number of potentially vulnerable devices in the billions, since almost all modern x86 PCs use UEFI firmware, and most of them trust the Microsoft Corporation UEFI CA 2011 certificate out of the box.

Revoking the shims

However, ESET reported its findings to CERT/CC and the vulnerable UEFI applications were all revoked.

The shims come from different tools such as PC diagnostic software, Linux distribution, and other UEFI-based utilities, the researchers explained. They also added that, since the attackers can bring their own vulnerable shims to any UEFI system with the Microsoft third-party UEFI certificate enrolled, they can exploit systems that are, at first, not affected.

To block the vulnerable shims, users should apply the latest UEFI revocations from Microsoft, it was said. While Windows systems will most likely do it automatically, Linux systems users should do it through the Linux Vendor Firmware Service.

“What makes these old shims dangerous is not a novel vulnerability; it’s that no new vulnerability is needed to bypass UEFI Secure Boot,” says ESET researcher Martin Smolár, who discovered the vulnerable shims.

“An attacker needs no complicated exploitation primitives — only a copy of an old, still-trusted but unrevoked shim binary and a basic understanding of how UEFI shims work. That is enough to bypass such an essential security feature as UEFI Secure Boot."

• Microsoft’s July 2026 Patch Tuesday fixed a record 622 vulnerabilities, including 58 critical, two exploited in the wild, and one publicly disclosed, plus 428 Chromium bugs
• Actively abused flaws include CVE‑2026‑56155 (AD FS privilege escalation) and CVE‑2026‑56164 (SharePoint privilege escalation), alongside notable issues in BitLocker and Copilot
• Surge in fixes is linked to Microsoft’s use of Anthropic’s Mythos AI, with patch volumes rising sharply since its adoption

Microsoft has released its July 2026 Patch Tuesday download, marking another record-breaking update, addressing hundreds of flaws across the ecosystem.

The release, which is currently rolling out to Microsoft users, fixes a staggering 622 vulnerabilities, including 58 critical-severity ones, two that were observed as being abused in the wild, and one which has already been publicly disclosed.

On top of that, Microsoft shipped fixes for another 428 Chromium bugs, as well.

A jump in numbers

There are simply too many vulnerabilities to mention all of them, however two that are being exploited in the wild are CVE-2026-56155 and CVE-2026-56164. The former is described as an “Insufficient granularity of access control in Active Directory Federation Services (AD FS)” bug, which allows an authorized attacker to elevate privileges locally. It carries a severity score of 7.8/10 (high).

The latter is a “Missing authentication for critical function in Microsoft Office SharePoint” bug that allows an unauthorized attacker to elevate privileges over a network. Microsoft assigned it a medium severity score (5.3/10), but the National Vulnerability Database gave it a 9.8/10 (critical).

Other notable mentions include CVE-2026-50661, a protection mechanism failure in Windows BitLocker that allows unauthorized attackers to bypass a security feature with a physical attack, and CVE-2026-48561, an improper neutralization of special elements used in a command in Microsoft Copilot, that allows an unauthorized attacker to execute code over a network.

If you think fixing 622 vulnerabilities in a month is a lot, you’re absolutely right. It’s well above what Microsoft is used to do, and this is most likely due to the company now using the fabled Mythos - Anthropic’s cybersecurity-oriented AI.

In June 2026, roughly a month and a half after the release of Mythos, Microsoft fixed 206 flaws, which raised eyebrows because it was significantly above the company’s usual amount of bugs fixed.

In May it fixed 120 flaws, in April 167, and in March - 79.

A new customer booking system, a bespoke CRM, or an app that can be sold as a new revenue stream. Small businesses are rarely short of great ideas, but often lack the time, funding, or technical expertise to go from idea to functional app.

This is the gap that vibe coding aims to fill, helping small businesses launch apps by combining plain English prompts with AI.

I caught up with Yoav Orlev, Head of Product at Base44, to get some insight into how small businesses can get the most out of vibe coding platforms. We also discuss some of the common pitfalls and risks associated with vibe coding and how to overcome them.

Vibe Coding has a lot of buzz around it, but what does it actually mean for the small businesses that can benefit from it?

Instead of needing to know how to build an application, small business owners can just describe what they want and it gets built. They’re essentially having a conversation with AI instead of wrestling with tools, templates, or code. There are significant advantages to vibe-coding, including:

Removing the “gatekeepers"

Previously, a small business owner who wanted a professional website had two options: pay a developer (expensive) or spend hours learning code (time-consuming). Vibe coding collapses both into a single conversation. A florist, a personal trainer, a local accountant, anyone can now describe their business and get a professional application without any technical knowledge.

Speed to market changes everything

A small business can go from idea to functional application in minutes, not weeks. For a new business owner, that's the difference between launching now and launching never.

Iteration becomes effortless

Want to change your text? Swap your color scheme? Add an agent? Instead of digging through menus or calling a developer, you just type what you want changed via natural conversion. Small businesses can now move and adapt as fast as their ideas do.

The playing field levels out

Enterprise businesses have had dev teams, agencies, and big wallets for years. Vibe coding gives small businesses access to the same quality output without the same budget. This is a significant shift, letting smaller businesses compete with larger ones.

Vibe coding means that for the first time, the barrier to having a custom application is no longer technical skill or budget. That's a meaningful unlock for millions of small business owners who previously felt locked out of the growing digital economy.

For years, small businesses have benefited from drag-and-drop tools. How has vibe coding changed this approach? What new benefits does it bring to the table?

Drag-and-drop was a genuine breakthrough. It opened up web creation to anyone, regardless of technical skill or budget constraints. Users needed to know what they wanted, where to put it, and how to make it look right, but they had more access and guidance than ever before.

Vibe coding starts the create-and-build process with a conversation. Users describe their business and goals, and the platform builds around that. It's a fundamentally different kind of interaction than drag-and-drop. Users are now not operating a tool but collaborating with one.

Vibe coding isn't here to replace drag-and-drop, and that's an important distinction. Both have real advantages. Vibe coding wins on speed and ease of use. Drag-and-drop gives users precision and hands-on creative control. The real breakthrough for small businesses is having both working together seamlessly by describing what they need, letting AI get there fast, then fine-tuning the details manually without ever switching platforms or starting over.

That combination is what gives small businesses a truly holistic way to build and grow an online presence. Less time building and more time focusing on their business.

If an entrepreneur wanted to build a tool for their business today, where should they start? What does a good prompt look like?

The best place to start is with the problem, not the solution.

Before opening any tool or writing a single prompt, an entrepreneur should get specific about what's actually slowing their business down. Is it taking bookings manually over the phone? Chasing invoices? Answering the same customer questions over and over? The clearest prompts come from the clearest problems.

When it comes to writing a good prompt, specificity is everything.

A weak prompt sounds like "build me a website for my business." A strong prompt sounds like "I run a mobile dog grooming service in Chicago with three employees. I need a way for customers to book appointments online, see my pricing by dog size, and get automatic confirmation texts."

The more context users give, such as industry, customer, specific workflow, and the outcome that’s trying to be achieved, the more useful the result. Think of it less like a search query and more like briefing a new hire on their first day. The AI works best when it understands not just what you want, but why you need it.

From there, iteration is your best friend. A first prompt doesn't need to be perfect. Get something on the screen, react to it, and refine from there. The most effective builders treat it as a back-and-forth conversation rather than a one-shot request.

What are some common examples of small businesses using vibe coding to build internal tools? Which do you believe are the most valuable and why?

The range of tools small businesses are building with vibe coding is remarkable. We're seeing everything from custom booking and scheduling systems, to lightweight CRM and lead tracking tools, client-facing portals, invoice generators, staff onboarding wikis, loyalty program trackers, and more.

The most valuable apps tend to be the ones replacing a manual process that was quietly costing the business time or customers. Custom booking systems are one standout. For service-based businesses, time is the product, and a system built around their exact workflow has a direct and immediate impact on revenue. Lightweight CRM tools are another reason because most small businesses aren't losing customers due to bad service; they're losing them because follow-up falls through the cracks. A tool built around how they actually sell, rather than how generic software thinks they should, makes all the difference.

The common thread is fit. What makes vibe coding genuinely powerful for small businesses is that they can build exactly what they need and tailored for them, rather than settling for something “close enough.” That's a shift that levels the playing field in a very real way.

Can entrepreneurs also use vibe coding to build customer-facing tools and new revenue streams? What are some good examples of this?

Absolutely! This where vibe coding starts to look less like a productivity tool and more like a genuine business accelerator. The same technology that helps a small business automate internal processes can also help them build entirely new products, services, and revenue streams that simply weren't accessible before without a development budget.

The examples are wide-ranging. A personal trainer can go beyond selling sessions and build a branded fitness app where clients track workouts and access custom programs, turning a one-to-one service into a scalable product. A marketing consultant can build a self-serve audit tool that generates leads while they sleep. A local chef can launch a meal planning subscription with a custom interface rather than relying on a third-party platform that takes a cut of every transaction. A retailer can build a personalized product recommendation quiz that increases average order value without touching their core website.

What all of these have in common is that they were previously only realistic for businesses with developer resources and big pockets. Vibe coding changes that entirely. The barrier to launching a new revenue stream is no longer technical and costly. It's just having the idea and the ambition to act on it.

Some business owners (myself included) have found ourselves stuck in a loop where AI fixes one issue, but breaks several others. What are some golden rules users can apply to get back on track when the ‘vibe’ goes wrong?

1. If possible, try to break down your request into smaller pieces. Sometimes, when trying to add too many features at once, the agent makes a mistake.
2. If something doesn't work at first, be very specific and clear. In many cases, it can also be helpful to state what the AI shouldn't do. For example, tell AI to change the layout, but make sure not to change any colors in the app.
3. Almost all vibe coding platforms allow users to revert. So if you are failing to add something 3 or 4 times in a row, then revert to the last successful version and try in a slightly different way.
4. Search the tool documentation. There are usually many examples and “how tos” that can carry you over a certain hurdle.

What are some of the pitfalls and risks that entrepreneurs should be aware of when vibe coding internal tools and potential products? 

There are 2 main pitfalls we see.

1. Not giving your product to users. We talk to a lot of users who keep building more and more features, since it's so easy, yet they hold off releasing the product, fearing it's too early. Don't. AI is all about moving as fast as possible. If you have a working product that will help someone, put it in the hands of users and get feedback as quickly as possible. Worry less about the perfect product and more about usage.
2. Security is huge. Most tools nowadays offer security scans, but as a builder, make sure you take the time to really understand how the scan works, what it covers, and what issues you might face. Most tools will cover you, but only if you know how to leverage them. Take the time to learn it or consult someone who can walk you through it.

When a trusted friend recommends a product or service to you, it can alter the way you think about the potential purchase. You tend to lower your guard, with the knowledge that some of the filtering work has already been done on your behalf.

This is particularly true when that friend understands your tastes and priorities, and how much you’re looking to spend; the recommendation carries weight before you even visit the website or look at the product page. By the time you click, a large part of the decision has already been made.

Now there is evidence that AI recommendations can have a similar effect on consumers.

We recently analyzed web traffic and conversion data from more than 35,000 ecommerce brands using Shopify and it revealed something significant: referrals from AI tools such as ChatGPT are converting at an average rate of 3.6%, compared with 1.23% for traditional Google search traffic. AI referrals are also generating around 30% higher revenue per session.

For ecommerce businesses, these findings point to an important change in consumer behavior that presents both an opportunity and a challenge.

The buying journey is changing

For years, digital marketing strategies have been built around search engines. A user enters keywords, compares links, visits multiple sites and eventually makes a decision. Brands compete to win the click, then persuade the customer after they arrive.

AI is changing the sequence.

Consumers are no longer simply typing broad phrases like “best running shoes” or “cheap Bluetooth speaker”. Instead, they are asking the likes of ChatGPT highly detailed questions tailored to their exact needs.

Where previously the user might have received a list of product websites to explore, now they are being given a specific tailor-made recommendation.

In effect, AI platforms are compressing the consideration stage of the buying journey. Much of the evaluation happens before the consumer reaches the brand’s website. By the time they click through, they already have a degree of confidence in the recommendation they have been given – much like the case of the recommendation from a friend.

This creates a different type of visitor altogether.

Traditional search traffic can often be broad and exploratory. AI-referred visitors, however, increasingly resemble ‘pre-qualified’ leads. They arrive with clearer expectations and stronger buying intent, which helps explain the higher conversion rates.

Why visibility now means something different

For brands, the opportunity is clear: higher intent traffic generally means stronger revenue per visitor and more efficient acquisition.

The challenge, however, is that many ecommerce businesses are still measuring success using frameworks built for the ‘traditional’ approach.

A large proportion of marketing strategies remain heavily focused on traffic volumes, click-through rates and keyword rankings. Yet AI recommendation systems rely on different signals. Visibility inside AI-generated answers depends less on traditional advertising tactics and more on credibility, authority and contextual relevance across the wider web.

Brands are no longer only competing for search rankings. They are competing to become trusted sources within the information ecosystem AI tools rely upon.

That has major implications for how companies think about content and discoverability.

Reviews become more influential because AI systems frequently incorporate them into recommendations. Third-party editorial coverage matters more because it contributes to authority and trustworthiness. Community discussions on forums and social media also gain importance because they help establish credibility and context.

The businesses that perform well in AI-driven discovery are likely to be those with strong reputations spread consistently across multiple trusted sources, rather than those relying purely on aggressive performance marketing.

Rethinking marketing measurement

There is another important lesson in this data. Many businesses may currently have an incomplete understanding of which channels are genuinely driving growth.

One of the perennial challenges with digital advertising is ‘attribution distortion’. In simple terms, platforms optimize for conversions, often retargeting existing customers or users who were already close to purchasing. This can create inflated perceptions of acquisition performance.

At the same time, AI-referred traffic may still be under-measured inside many organizations because it remains a comparatively new source of inbound visitors.

The danger is that businesses continue over-investing in channels that appear successful according to legacy metrics while underestimating emerging forms of high-intent traffic.

This matters because AI-driven discovery is likely to become more important over time.

Consumer expectations are evolving rapidly. People increasingly want answers tailored to their exact circumstances rather than broad lists of generic options. AI interfaces are naturally suited to that type of interaction because they can process nuance and context in ways traditional search engines struggle to replicate.

What brands should do next

Brands therefore need to rethink not only where they advertise, but how they present their offer online altogether.

That starts with understanding how the business appears across the wider digital landscape. Are reviews consistent and trustworthy? Is the brand being referenced by credible publications and communities? Is product information clear, accurate and useful?

It also means creating content that answers real consumer questions in detailed and genuinely helpful ways, rather than simply targeting high-volume keywords.

None of this means traditional search is disappearing overnight. Search engines remain hugely important, and publishers continue to play a central role in shaping the information AI systems consume and reference.

What is changing is the path consumers take before making decisions.

The era of winning attention purely through visibility is gradually giving way to an era of winning trust before the click ever happens.

We've tested the best email marketing platforms on the market.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

AI took enterprises by storm, with many opting for integration as fast as possible in fear of falling behind the more ambitious tech adopters. But speed alone isn’t always an advantage and as a result, 95% of enterprise pilot programs still failed to deliver measurable financial returns just last year.

Now that we’re a few years past the initial AI explosion, the pressure is on to prove true ROI from these projects.

Businesses have most frequently poured resources into AI tools aimed at boosting productivity and automating workflow in hopes to target the most universal, longest standing business goal: doing more with less. But what leaders should be doing is identifying where the technology can solve the biggest issues specific to today’s business climate. One of those issues is cash flow forecasting.

This year, 52% of American CFOs named cost management as their most worrisome internal concern. While a well-oiled cost management strategy remains critical for creating a strong financial cushion and remaining resilient, balancing fixed operations with constantly shifting real world variables is never easy.

As companies increasingly look for ways to remain nimble and improve decision-making, those that can leverage AI to forecast trends in cash flow demand, churn risk, and spending pattern shifts will find themselves on a quicker path to ROI. Doing so requires harnessing the right data, and these financial signals are hidden in the transaction layer.

While businesses have long mined transaction data for traditional analytics and reporting, it’s far under-utilized in AI strategies. There needs to be a shift from viewing these insights as archival records of past performance to real time indicators of what’s to come.

Making revenue forecasting more adaptive

There are several revenue indicators that lie within bottom funnel operations that AI has the ability of turning into actionable insights. From frequently adjusted terms within contract renewals to the average time it’s taking customers to finalize transactions, purchase signals like these can help AI systems make smarter predictions about demand or accounts receivable.

To provide a more granular view into the value of this data layer, let’s look at upgrade or renewal activity for example. Customer retention is a key element to maintaining predictable cash flow and is among the first to go during an economic shakeup.

Tracking accounts that consistently upgrade a product or service on time to see that they suddenly miss a milestone could immediately flag eventual churn risk. These deviations should also be compared across similar accounts to segment risk based on geography, product lines, or size and industry.

From there, leaders can act proactively with strategies like targeted discounts or incentive measures to encourage retention. Alternatively, accounts that are expanding faster than expected could provide predictions into other customers who might be ready for higher value offerings.

Identifying cues like those that often precede cancellations, along with delayed payments, reduced usage, or smaller order size for instance, can equip finance or leadership teams with rolling forecasts. Whereas on the other hand, monthly or quarterly forecasts typically only rely on historical averages and don’t provide the real time guidance needed for a quickly shifting marketplace.

This can manifest into a powerful decision-making engine. One that is dynamic enough to support flexible cost management strategies. Seeing where cash-flow is moving allows leaders to make more informed decisions.

For example, if it’s consistently being found that these customers are upgrading at slower rates then it may be a good indicator to preemptively reduce inventory or relax timelines for product development teams. In turn, leaders can avoid allocating too many resources to demand that may not end up materializing.

This also gives teams more flexibility to adjust spending and production before any cash flow pressure sets in.

An important element to keep in mind is that these purchasing behavior insights often sit across separate systems. While sales teams may have insights into average order values, only legal or finance may know how payment terms are changing across clients.

First mapping where all of these metrics currently live is critical to then unify them into one place for predictive models to cross-analyze everything against each other and make stronger recommendations.

Connecting to ROI directly

Many businesses have revolved their AI projects around generative AI for efficiency gains in producing content or developing software for instance. But not only are these task-level initiatives harder to prove a measurable impact from, in some cases they end up hurting productivity in the long run with added time spent reviewing and editing AI outputs.

When it comes to analyzing data for forecasting and predictions, AI has shown immense value and is tied to more tangible business outcomes. Now, forecasting real time insights tied directly to revenue can help leaders remain adaptable in an increasingly unpredictable economy, offering a fast track to ROI on these projects.

We've featured the best AI website builders.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• New York State has temporarily banned new, large-scale data centers
• One-year moratorium affects new 50MW+ campuses
• NY wants to assess community, environmental and grid impacts

New York State has banned the construction of new hyperscaler data centers, marking the latest step as mounting local opposition against AI and cloud facilities builds, with communities citing concerns over rising electricity costs, water consumption and other environmental factors,

State Governor Kathy Hochul is responsible for introducing the first statewide moratorium on new campuses, which imposes a one-year pause while the state looks into the environment, energy supply and communities.

Projects that have already been permitted will still continue as expected, but new buildouts will face restrictions and delays until the moratorium is lifted.

Data centers banned in NY - for one year

The one-year ban buys the state time to evolve regulations to address some of the challenges presented by large hyperscaler data centers, including strained grid supplies, emissions and other environmental impacts, and stresses on local communities.

Under the new ban, campuses that require at least 50MW of electricity will be affected, which for an AI data center isn't all that much. Some of the largest measure power consumption in the hundreds of megawatts, or even gigawatts in the case of high-profile, flagship campuses like OpenAI's Stargate Project.

"[The moratorium] comes as the direct result of immense public pressure from people across the state demanding their elected leaders protect them from Big Tech's assault, which threatens the state's clean air and water and New Yorkers' financial security," New York State's Food & Water Watch Director Laura Shindell said.

"New York has always been at the forefront of innovation and change but we’ve also always guaranteed that New Yorkers benefit," Hochul noted.

Although New York represents the first statewide temporary ban, other regions have also been looking to pause buildouts as they assess the damages. Just last month, Seattle also voted to ban new projects for a year.

For years, the central question around AI was whether it really works the way much of the market’s hype says it does.

That debate is settled, and we’ve seen what AI can do. The more important question now – and one most organizations still aren’t asking – is whether we’re working alongside it the right way.

That shift in framing matters. Budgets have been allocated, tools deployed, and pilot programs have graduated into full-scale production.

And yet, something is still missing.

IBM’s Institute for Business Value found that only 25% of AI initiatives have delivered their expected ROI, and just 16% have successfully scaled across the business despite years of investment and genuine enthusiasm for what the technology can do.

The problem isn’t AI itself; the bottleneck stems from what organizations have or have not built around it.

That’s a leadership problem, and fixing it will require more than buying better tools or scheduling more trainings.

Stop Investing in the Wrong Places

The instinct for most organizations has been to buy the latest platforms, stand up a few pilot programs, and bring in a vendor to train their workers. That approach addresses the surface-level challenge, but misses the greater underlying issue. The truth is the greatest barrier to AI maturity is the lack of investment in the human infrastructure needed to support it.

The companies seeing the strongest AI outcomes are rarely those with the most sophisticated or expensive models. They’re the ones that have fundamentally rethought how their people work. Among organizations that Boston Consulting Group designated as AI leaders, roughly 70% of resources went towards people and process changes, 20% to IT infrastructure, and only 10% to the AI models themselves. Most organizations have that ratio backwards.

When leaders become hyper-focused on deploying the right tools and launching the right uses cases, they neglect the organizational muscles that are essential to using AI responsibly and consistently. All the tools in the world won’t close that gap without the right training, guardrails, and policies to back them up. And building that support structure must be a leadership priority, not an afterthought.

The Productivity Gains Are Real, But Fragile

None of this is to say AI isn’t creating real value. It absolutely is – at least, for the companies using it well. But those gains are more fragile that many leaders realize. They evaporate when companies lack support for their employees across their interactions with the technology, or when they fail to clearly communicate where human judgement and critical thinking are still essential.

The data here is hard to ignore. Among employees who use AI on the job, less than 8% report receiving extensive training with their tools. And that number has barely budged despite a sharp increase in daily usage. Moreover, 60% say it often takes them longer to figure out how to accomplish a task with an AI tool than it does to simply do it themselves.

Companies are deploying AI faster than they are enabling people to use it, and in doing so, may be creating exactly the friction and confusion they were trying to eliminate.

What Leaders Owe Their People

This is where leadership has to show up differently. The gap between AI potential and AI reality isn’t going to close through procurement decisions or new rollout announcements. It closes with deliberate, ongoing investment in people. That means three things:

Focus trainings on people, not just tools: AI is evolving faster than training curriculums can keep up. Invest in training role-specific judgements, helping people understand where AI makes them faster and where it introduces challenges or risks.

Move beyond adoption rate metrics: If 80% of your organization is using AI and productivity is still flat or declining, adoption is the wrong metric. Measure time-to-completion on real work tasks and be honest about what you find. Some use cases that are slowing people down simply shouldn’t be using AI.

Stop treating AI policy as a compliance checkbox: Companies getting this right have built AI governance into how they plan and execute work daily, not appended into and acceptable use document. That means leaders who model where they use AI and where they don’t, and who are willing to explain why.

Making AI Potential a Reality

The technology is ready. But leaders need to be able to do more than allocate budget and monitor usage.

They need to decide when AI should and should not be used, what to rebuild rather than automate, and how to support their teams throughout all of it.

Those are the questions most organizations are still failing to ask, and until they do, the ROI gap isn’t going anywhere.

We've ranked the best HR software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• SAS finds EU SMBs are seeing better results from real-world AI deployments
• Compliance is still a major headache for one in four leaders
• Preparing for the EU AI Act could have helped

A new AI readiness report from SAS has claimed European small businesses are among the most ready to achieve the most ROI from artificial intelligence compared with other global regions.

While North American SMBs perform better in planning, building and enabling, they fall short on actual deployment, indicating that European SMBs have progressed beyond pilots to actual implementation.

However, the number of businesses globally that are still in the earlier stages of AI readiness is much greater, with 37% considered 'experimental' and 33% considered 'opportunistic'. Nine in 10 of the stage-one (experimental) companies haven't even got a formal AI strategy in place, the report reveals.

AI deployment is higher in Europe, but overall deployment is still low

Compared with the number of businesses in the early stages of AI readiness, only 9% have fully embedded AI into strategy, operations and decision-making.

For many, it's the fundamental building blocks that are still holding them back. Nearly half say data is still scattered across systems (45%) and that AI tools operate independently from one another (46%). Compliance, security and risk management is still the biggest barrier for 24% of the survey's respondents.

"Organisations treating governance as a foundation rather than an obstacle are often the ones best positioned to execute," SAS Global Channels SVP John Carey wrote.

The report also implies that preparing for regulation like the EU AI Act could have positioned companies better to maximize their AI ROI, while other regions are seemingly lacking in this area.

IDC Research VP Daniel-Zoe Jimenez urged businesses to focus on aligning data, people and resources to avoid disconnected pilots. "Experimenting with the technology is one thing. Deploying it strategically and sustainably is quite another."

• White House, Treasury, DHS and DoW come together to launch Gold Eagle scheme
• The initiative will prevent duplicated work and prioritize vulnerability remediation
• Gold Eagle will also help to identify which systems could be at risk

The US Government has launched Gold Eagle, a new clearinghouse which looks to centralize vulnerability discovery and remediation against a backdrop of evolving AI-powered security threats.

Gold Eagle will serve as a central hub between federal agencies, AI developers, open-source software developers and critical infrastructure companies, in a bid to increase the speed of vulnerability discovery and prevent major incidents from occurring in the first place.

The scheme came about under President Trump's June 2 2026 executive order 'Promoting Advanced Artificial Intelligence Innovation and Security' and represents collaboration between the Treasury, the DHS' Cybersecurity and Infrastructure Security Agency (CISA) and the Department of War.

US Gold Eagle scheme addresses growing vulnerability exploitations

Under the scheme, vulnerabilities scanning will happen centrally to ensure multiple organizations aren't independently repeating the same work. Gold Eagle will also identify which software, networks and critical infrastructure could be at risk, before coordinating fixes. The White House described the scheme as a "force multiplier."

Although AI is largely to blame for the increase in attacks, Gold Eagle is set to fight fire with fire by employing AI to identify bugs too, using models like Anthropic's Mythos.

"Through this strategic partnership, we will expand existing security measures to safeguard software and networks in the 21st century and continue to promote advancements in artificial intelligence," DHS Secretary Markwayne Mullin wrote.

The concept of a dedicated clearinghouse centralizes vulnerability management to ensure the right bugs are being prioritized and to cut through the noise of lower-quality reports. Its assistance will most likely be felt by the open-source community, which has limited resources and financial backing to identify and fix issues as effectively as enterprise software vendors.

"Under the leadership of President Trump, we are bringing a wartime footing to the cyber domain to relentlessly patch vulnerabilities," Secretary of War Pete Hegseth added.

The National Cyber Security Centre (NCSC)’s recent advisory detailing how the Russian cyber actor APT28’s exploited vulnerable routers to enable DNS hijacking highlights a dangerous blind spot in enterprise security today.

Traditional perimeter-based security models are increasingly failing at the edge, and this should be uncomfortable reading for any security leader still anchoring their strategy to perimeter defense.

This is the same story we saw play out in 2024, when Chinese state sponsored bad actors linked to Volt Typhoon exploited an unpatched, end-of-life FortiGate 300D firewall to compromise a domain admin account, escalate privileges, create a new user, and establish persistence deep enough to survive a device restart.

One unpatched edge device. One breach vector. Total access. The perimeter didn’t just fail; it handed attackers the keys. According to the FBI, this breach remained undetected for over 300 days, and the same exploit was used to gain access to over 100 separate utility companies across the US.

These two incidents aren’t outliers. They’re a pattern. And the pattern is telling us that the old model is finished. Add to this the recent news about the power and sophistication of the latest generation of agentic AI tools such as Anthropic’s Claude Mythos, and it really is time to act quickly.

The problem with perimeter thinking

Legacy security was built on a simple assumption: draw a line between trusted internal systems and untrusted external ones, defend the line, and you’re protected. It made sense when company data lived in on-premises data centers and employees showed up to the office. The boundary was real and defensible.

That world is gone. Edge nodes are now everywhere; in factories, retail stores, utility substations, and customer premises, and the clean network borders of the past have either blurred or disappeared entirely. Edge devices themselves rarely have dedicated security capabilities. IoT and OT systems in particular frequently lack the robust features needed to detect and resist advanced threats.

Worse, software-based management tends to fail precisely when it matters most. When the software layer is compromised or unresponsive, organizations lose visibility and control at the exact moment they can least afford to.

Then there’s credential theft. According to Verizon’s 2025 Data Breach Investigations Report, the human element is a factor in 60% of breaches and attackers have become highly effective at exploiting it. They don’t need to break in, they walk in, using legitimate credentials, through the front door.

Once inside, now often augmented by AI capabilities, they move laterally from a single-entry point across operational systems, compromising entire environments in minutes. A security architecture built around password authentication offers almost nothing against this type of attack.

What resilience actually looks like

The security conversation has shifted. For years the focus was on keeping attackers out. That is still necessary, but it’s no longer sufficient. What organizations are increasingly recognizing is that resilience depends just as much on what happens after a compromise; specifically, whether security teams retain visibility and control when it counts.

Lost visibility during an incident, even a contained breach, can quickly spiral and escalate. This is driving a serious rethink of how distributed infrastructure is managed, particularly as edge environments extend further across factories, retail locations, utilities and a myriad of remote sites.

Out-of-band management (OOBM) is one approach gaining real traction here. Rather than relying on the production network for management traffic, OOBM operates on an entirely separate, highly secure parallel path. That means that even when the main network is compromised or down, edge devices remain manageable, visible and controllable.

Crashed devices can be remotely rebooted, or even re-configured. Powered-off devices can still be reached. And critically, administrative access is separated from the primary production network targeted by attackers, thereby reducing exposure to the credential-based attacks that are now the most dominant breach vector.

The operational benefits are real too: fewer costly emergency site visits, faster data recovery times, and preserved control during high-pressure incidents when software management tools have gone dark.

The reckoning

What’s becoming clear is that perimeter security, on its own, is no longer a viable strategy. The edge has expanded too far, credentials are too easily stolen, and attackers are too fast once they’re in.

The organizations that will weather the next wave of incidents aren’t necessarily those with the most sophisticated perimeter defenses. They’re the ones that have accepted the compromise will happen and built the visibility, control and recovery capability to deal with it when it does.

Everyone else is building blind spots. And attackers, increasingly, know exactly where to look.

We've reviewed and ranked the best password managers.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

In February 2026, the California Attorney General announced a $2.75 million settlement with Disney DTC and ABC Enterprises, the largest in the history of the California Consumer Privacy Act (CCPA).

Most of the noise has focused on the legal and interpretive dimensions: how the settlement reads against the statute, what it signals about enforcement posture, and what it adds to an enforcement record that has been building for more than three years.

Since the first major CCPA action against Sephora, in 2022, for failing to honor opt-out requests, settlements with DoorDash, Tilting Point, Healthline, Sling TV and Jam City have progressively expanded the technical specificity of what compliance has to look like in practice.

For privacy leaders trying to decide what to do on Monday morning, the most useful lessons from the Disney settlement are technical. In many enterprises, there is a gap between two initiatives that operate in parallel: a privacy tech layer focused on cookies, cookie banners, and webforms, and a data layer that runs on identity graphs, pseudonymous profiles and cross-system data flows.

The four technical claims leveled by the California Attorney General: gaps in how opt-outs reach logged-out users, disconnected opt-out tooling, missing cross-brand propagation, and absent opt-out functionality on apps and connected TV, are four symptoms of this misalignment

These challenges and claims are not unique to Disney, nor do they reflect negligence. They are consequences of timing. The privacy layer in most enterprises is built to satisfy a regulatory model that took shape between 2018 and 2022, when "consent management" largely meant deciding which cookies fired on a webpage.

The data layer it was bolted onto had been evolving for a decade by then, towards exactly the kind of cross-device, partner-dependent identity resolution that makes modern advertising and personalization possible. The two were never properly integrated in most enterprise environments because, for several years no one seemed to mind.

What the recent enforcement record establishes is that they do now.

One principle, four symptoms

The settlement's central principle is unusually direct: "If a business can associate a consumer's devices with the consumer for advertising purposes, it can and must associate those devices with the consumer for purposes of honoring the consumer's opt-out rights."

In other words, the scope of your opt-out obligation is defined by the scope of your data monetization, not by the scope of your consent management platform. If your advertising stack resolves an anonymous device signal to a known profile to target a consumer, you have, for the purposes of the law, identified that consumer.

The obligation follows the use of identity capabilities, not who built them. If identity is being used to target a consumer with ads, the same identity must be used to exclude them once they opt out.

That single principle unifies the four technical points of the settlement.

The first: identity parity concerns whether opt-outs cover consumers who aren't currently logged in. Many enterprise programs apply opt-outs only to authenticated users, on the reasoning that the business doesn't know who a logged-out user is. The settlement reframes that. If your advertising infrastructure uses pseudonymous profiles—the device-level identifiers that ad-tech systems use to recognize the same person across visits without a login - to target that user, then for the purposes of the law, you have identified them. The opt-out has to reach the same identity.

The second: architectural fragmentation concerns how opt-out requests travel through the systems meant to enforce them. Most enterprise privacy programs run two separate products: a consent management platform (CMP) that governs which trackers and tags fire on the website, and a data subject rights (DSR) tool that handles webform submissions like Do Not Sell or Share requests. When those products aren't integrated, a consumer who submits a webform may stop appearing in certain backend records, but the CMP keeps firing the same tags on every page they visit, and data sharing continues. The webform captured the request. The collection infrastructure never received it.

The third: cross-brand propagation concerns whether an opt-out submitted on one property reaches the others that share its data infrastructure. If a media company runs three streaming services on a single advertising stack, and a customer opts out on one, the law treats that as an opt-out across all three because the data is monetized across all three. The technical capability to propagate the signal usually exists; the compliance logic that ties opt-outs to it rarely does. The same point extends downstream: if ad-tech partners already hold a consumer's data, blocking tags on your own site doesn't reach what they have. They need to be actively notified, through a workflow that runs every time an opt-out is processed.

The fourth: non-browser surfaces concerns the consumer-facing channels that aren't a website. Cookie-based consent tools were built for browsers. They do not, by default, extend to mobile apps, connected TV environments, or any other surface where data is collected outside the browser. In Disney's case, consumers on a connected TV app could only opt out by going to a webform on a different device - a webform that had no effect on the code transmitting data from the TV app to its ad-tech partners. The mechanism existed; the obligation it was meant to satisfy went unmet.

The difficulty is structural

The four technical gaps above share a common root, and fixing them points to something more fundamental than patching individual systems. None of this is straightforward to operationalize.

Modern data infrastructure is genuinely complex and rebuilding how the privacy and data layers of an enterprise stack actually communicate is not the kind of project that gets completed in a quarter. But the structural difficulty points to something deeper than implementation effort: a flaw in how most organizations have framed the problem from the start.

Privacy as a data infrastructure question, not a regulatory one

Privacy programs built around regulations are, by design, reactive. A new law passes, a settlement lands, an enforcement sweep reveals an unexpected gap, and the program scrambles to respond. That cycle is the predictable outcome of treating privacy as a compliance checklist rather than a capability embedded in how data actually moves through the organization. Regulations will keep coming, and they will keep changing. No program designed around any single regulatory framework will stay current for long.

The more durable approach starts in the data infrastructure itself. When privacy controls are built into the data layer, tied to identity resolution and data flows rather than bolted onto individual regulatory requirements, they adapt. A new regulation adds a specific obligation, but the underlying framework for honoring consumer choices across identifiers and systems is already in place to absorb it. The work becomes configuration, not reconstruction.

The path forward is not another tool layered on top of existing infrastructure to satisfy the next regulation. It starts by embedding privacy controls into the data layer itself, tied to how data actually flows through the organization. That foundation does not need to be rebuilt every time the regulatory landscape shifts. It absorbs change. For privacy leaders looking to get off the reactive cycle for good, that is where the work begins.

The best business VPN helps you connect safely and securely online to your business.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The disclosure-to-exploit window used to be measured in weeks; for weaponized vulnerabilities, it now runs in hours. Out-of-cycle patches that used to be exceptional have become routine across enterprise environments of meaningful size.

This pattern now has a name. You might have heard it already: the Patch Apocalypse. Sounds a bit dramatic, but the impact warrants the drama.

It describes something measurable — software flaws are being discovered, disclosed and weaponized faster than most patch management programs were built to handle.

Several factors are converging at once. Frontier AI models are accelerating vulnerability research — Anthropic's Project Glasswing and comparable initiatives have produced thousands of high-severity findings in compressed timeframes.

Attackers are using the same class of tooling to reverse-engineer patches far faster than previously thought possible. Public disclosures are arriving on shorter cycles.

For any team responsible for keeping production systems patched, this all translates to a backlog that grows faster than the available maintenance management windows can drain it. And “drain” is the right word here, because that’s also the impact on the team: it’s very, very draining.

This is far from an anecdotal observation. The workforce cost is already visible.

Recent UK data shows what’s happening at the personnel level: 42% of UK IT professionals report high levels of stress from their jobs, and 76% say that stress is affecting their physical and mental health. 30% report difficulty concentrating, 35% report trouble sleeping and 30% report increased anxiety and depression.

Why traditional patching is breaking down

Patch management was built around predictability. Vendor releases on a known schedule. A defined maintenance window. Manual testing in a staging environment. Communication, approval, deployment, verification.

The model worked when most enterprise software was released on predictable monthly or quarterly schedules, when threat actors needed weeks to weaponize a disclosed CVE, and when out-of-cycle patches were rare enough that a program could absorb them without restructuring. When those scenarios change, the model’s validity changes.

Two structural changes have done most of the work:

Volume is the first. When a single AI model can autonomously surface thousands of high-severity vulnerabilities across major operating systems and browsers — as Project Glasswing did within weeks of its April launch — the downstream effect is more CVEs arriving sooner, with public patches available, all flowing into the same backlog the IT team was already trying to clear.

Velocity has compounded that. Attackers have access to the same class of capability. Patches can be reverse-engineered in as little as 72 hours, sometimes far less. Any system unpatched in that window is exposed to working exploits.

Combine the two, and a patch program that was already running near capacity has to absorb a step change in volume, with shorter deadlines and less predictability about when the next critical disclosure will land. That’s a lot of pressure, and it’s what’s driving the stress data up.

Automation is taking center stage

As an operating model, automation is far better-equipped to survive a Patch Apocalypse than previous iterations. Three important principles underscore the model’s efficacy:

1. Continuous, risk-based triage. The CISA Known Exploited Vulnerabilities list is the non-negotiable top tier. An Exploit Prediction Scoring System threshold appropriate to the environment can drive priority for everything else. Below that threshold, work waits for the maintenance ring.

2. Automated test and deployment rings. The test cycle has to compress to fit the exploit window. Even with top-tier skills and best intentions, a human checklist cannot move at that speed. The familiar sequence — test ring, early-adopter ring, broad production, mission-critical — has to be instrumented and capable of running without manual coordination at every stage.

3. Closed-loop verification. A patch isn’t deployed until the install is confirmed on every endpoint, and a CVE isn’t closed until a rescan confirms it. Compliance evidence is produced as a byproduct of the workflow, not assembled from a spreadsheet the week before an audit.

Industry data points in the same direction. 67% of IT professionals say AI tools and automation will free up their time for more interesting, fulfilling work. 66% say the same tools will help them provide better service to end users. Less than one in three organizations report having fully embedded automation in their IT workflows.

Who’s paying the cost

Any cost considerations regarding patch programs need to take the human cost into account. A patch program running on legacy assumptions will absorb the Patch Apocalypse by burning out the team running it. The stress figures are showing the early signs. The downstream cost includes attrition, error rates, lost productivity and the slow erosion of the institutional knowledge that holds a program together.

On the other hand, programs where automation runs the workflow have the potential to absorb the same volume without requiring the team to absorb it personally. Continuous prioritization, instrumented rings and verification embedded in the workflow take variable, manual work out of the critical path.

Two-thirds of IT professionals see AI and automation as a route to better work — fewer frantic escalations and more time on the problems that need human judgement.

The Patch Apocalypse is very much here, and is poised to reach every program. Is the workflow underneath built to absorb the impact? If not, consider the whole scope of what — and who — is at stake.

We've reviewed and ranked the best endpoint protection software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• UN Secretary General calls for a global ban on autonomous "killer robots”
• Guterres argues that delegating life-or-death decisions to machines is “morally repugnant”
• Governments should take a stance now – not wait for something catastrophic to happen

UN Secretary General António Guterres has called for lethal autonomous weapons, which he describes as ‘killer robots’ to be prohibited under international law following recent discussions at the first Global Dialogue on Artificial Intelligence Governance in Geneva.

Guterres’ demand to ban these weapons focuses on those capable of identifying, selecting and attacking targets without human oversight, which leaves artificial intelligence and other computer systems in charge of a life-or-death decision.

He ultimately argued that certain decisions must remain exclusively human, and the decision to take a life is well into the boundary of requiring human oversight. Transferring the decision-making to killer robots would be “morally repugnant” and “politically unacceptable,” he argued.

AI requires global regulation as military AI poses major threats

Key to the Secretary General’s argument is that he urges governments to take action and ban such robots now, rather than waiting for an autonomous weapon to cause a major incident before rethinking their strategies.

“Let us not wait for atrocity to act,” Guterres said. “Some decisions must remain forever human – none more than taking a human life.”

The issue is becoming more urgent now that AI models and advanced chips are already being used within military intelligence, targeting and other battlefield systems.

More broadly, Guterres’ thoughts align with those of Anthropic, which recently had a dispute with the Pentagon after seeking guarantees that its models would not be used for autonomous weapons or surveillance.

While the Pentagon had rejected those limitations, arguing that it should be able to use Anthropic’s models for any lawful purpose, the case highlights how private companies are becoming increasingly intertwined with digital warfare.

Reporting by the Wall Street Journal cited a similar view by Pope Leo XIV, who warns that AI-controlled weapons could promote an “anti-human” view of warfare. He warned that the autonomy could reduce some dangers and distance political leaders from the human consequences of conflict.

There’s a need to balance the pros and cons of AI

However, artificial intelligence does promise several benefits to modern warfare, particularly in its ability to process huge amounts of information extremely quickly. With modern compute, militaries can respond to threats at lightning speed, improve their accuracy and precision, reduce soldier risk and potentially reduce civilian casualties, too.

Critics also question whether human oversight of AI systems is at all meaningful if the person in charge only has seconds to act on AI-generated information in the first place.

It’s also yet to be determined which party or group of parties should be held accountable for any incidents or mishaps – human operators, commanders, hardware manufacturers and software developers are just some of the parties up for judgment.

“We may be the last generation able to set the terms on which humanity and machines coexist,” Guterres warned separately in an X post, warning that AI must be governed, trusted and fair.

“It sounds like science fiction, but it's a real possibility, and it could change the world in ways that we don't understand yet, and it could change the power dynamics of our planet in ways that require our attention,” Independent International Scientific Panel on AI Co-Chair Yoshua Bengio added.

• Kokuyo Energy Line replaces discrete sockets with one continuous slot that takes up to five two-prong plugs anywhere along its length
• It won a 2025 Good Design Award thanks to its accessibility, style, and attention to detail
• The Kokuyo Energy Line also offers an inclusive design that allows users with upper limb disabilities to easily use the product with just 1 hand

Kokuyo is a well-known Japanese furniture giant that focuses on both functionality and aesthetics across its furniture and interior design lines.

It regularly collects Good Design Awards even as it runs a public "live office" where users can test out hardware in a workplace setting, in addition to HOWS DESIGN, an inclusive-design program that has birthed multiple successes.

The Kokuyo Energy Line power strip is one such offering: it comes with a Good Design Award (2025) and is influenced by the HOWS DESIGN program.

A power strip that blends focus on design, functionality

Modern work desks are a far cry from their older, mundane alternatives at the workplace or at home, offering a degree of customizability that few could have foreseen.

Amid all the upgrades one sees in monitor arms, cable management, electric desks, and even headphone stands, one thing often remains an aesthetic (and often functional) outlier: the modern power strip.

The Kokuyo Energy Line aims to fix this by offering a single continuous power socket that can house up to 5 appliances, with a plug-anywhere design and wiring concealed under the table via a minimalist clamp.

It also leverages an inclusive design that allows users with limb disabilities to use it easily with one hand, as demonstrated during the company's inclusive design workshop at its "HOWS PARK" diversity office.

The Kokuyo Energy Line clocks in at about 7,000 Yen (~$50) and comes in both black and white colors to suit different themes. It aims to eliminate cable clutter on one's desk with a one-size-fits-all approach.

Despite the advances made, the Kokuyo Energy Line also comes with its limitations: the design means it does not accommodate heavy-duty appliances, with a 5-device limit and a 1500W power ceiling, which may leave users who use high-end computers or multi-monitor setups looking for something different.

Unlike competing power strips, it also lacks a surge protector or grounding support, which limits its use to electronics that use a 2-prong cable.

With no support for 230V power cables or sockets and no plans currently to incorporate them, the Kokuyo Energy Line is an impressive but geographically limited power strip offering that suits the company's target audience looking for an aesthetic upgrade, albeit with serious limitations that could see it not be in play for more demanding consumers in terms of sockets and/or power.

• Samsung's PM1763 entered mass production as the fastest SSD on paper, focusing solely on AI data centers as its key market
• The PM1763 offers read and write speeds of 28.4GB/s and 21.9GB/s respectively, essentially twice that of its predecessor, the PM1753
• The drive can't physically be used in consumer-grade PCs, adhering to an EDSFF-only form factor while also requiring PCI-E 6.0 channels, something that has yet to be available to end-users

Samsung has announced it is now mass-producing the PM1763 SSD, which aims to replace the PM1753 as its highest-end enterprise-class SSD for AI customers.

The PM1763 offers read speeds of 28,400 MB/s and write speeds of 21,900 MB/s, leveraging PCIe 6.0 connections.

It uses the company's 9th-generation V-NAND, along with a 4nm controller, to deliver these speeds even as PCI-E 6.0 offers double the per-lane bandwidth available to users.

A very fast SSD that narrowly beats the competition where it matters

Samsung's offering is, at the time of writing, without doubt, the fastest SSD available to enterprise clients on paper, but it does have a few caveats.

The company claims the PM1763 offers "industry-leading performance", and that is definitely true in both the read and write departments, especially the latter, but it barely ekes out a win in the former over the Micron 9650.

The Micron 9650 offers read speeds of 28,000 MB/s and much slower write speeds of 14,000 MB/s sequentially, also leveraging PCI-E 6.0 to deliver such performance.

Samsung's SSD is decisively faster on another metric that is key for AI customers, however: it offers 6.92 MIOPS in sequential read speed versus Micron's 5.5 MIOPS.

Micron's offering, however, has already been in mass production since February 2026 and is expected to enjoy greater availability for the rest of the year than Samsung's enterprise flagship.

Samsung's offering also incorporates other gains: it delivers power efficiency that the semiconductor giant says is 1.8x better than the PM1753 and supports both post-quantum cryptography (PQC) algorithms and the TEE Device Interface Security Protocol (TDISP).

It must be noted that both Micron and Samsung's offerings are only part of the puzzle, as enterprise consumers are currently gearing up for the next generation of server hardware. Both Nvidia's Vera platform and AMD's EPYC "Venice" offer PCI-E 6.0 connectivity that these drives need to run at maximum speeds.

One would expect similar gains soon in the consumer market, where Samsung's Gen 5-based 9100 Pro is one of the few that currently rule the roost with advertised read and write speeds of 14,800 MB/s and 13,400 MB/s, respectively, but that might be wishful thinking at best.

The gains from the PM1763 are not expected to trickle down to consumers for a multitude of reasons. Primarily, PCI-E 6.0-supporting hardware does not currently exist at the consumer end, even as datacenters begin to adopt it.

The bleeding-edge storage on offer is also expected to be prohibitively expensive, pitting consumers against datacenter clients with seemingly limitless pockets for now, and industry figures such as Phison's CEO are already warning that AI demand will keep NAND and DRAM in shortage through 2026; consumer storage is increasingly built from what the data centers do not take.

The PM1763, therefore, at least from an end-user's perspective, might as well be a proof-of-concept SSD; it is unlikely to make its way onto their desktop anytime soon, and they are unlikely to be able to afford it unless they want to host a data-center-class server at home.

The Gen 6 storage era has arrived, attached to hardware you cannot buy, in a shape you cannot mount, on an interface you do not have, built from NAND that was never going to reach you anyway. Sadly for enthusiasts looking for a faster SSD: The speeds are real. So is the velvet rope.

For as long as there's been software there has been tension between open source and closed source. The creator of the Linux kernel, software engineer Linus Torvalds, has long been an outspoken advocate for open source environments – while putting the boot into closed source corporate entities who have tried to exploit the ecosystem.

Open for business

During a Q&A at Aalto University, an attendee complained that the Nvidia Optimus chip – a processor that switches between integrated graphics and dedicated graphics – was no longer being supported on Linux.

That meant that this expensive chip fitted into her Linux laptop was basically useless unless developers worked on a fix by reverse-engineering the process. Torvalds answered by giving Nvidia both barrels – complaining that the company was profiting from the Linux Foundation ecosystem through products like the Tegra chips for Android devices.

At the time, he complained that competitors like Intel and AMD were happy to co-operate with Linux to build native drivers whereas Nvidia would treat its proprietary code as classified material. This exchange culminated with an angry Torvalds flipping Nvidia off to camera.

Turning the tables

The situation today is almost entirely different from that we found nearly 15 years ago. Nvidia has embraced open source with open arms, and has also fully transitioned to open source Linux GPU kernel modules as being the default options as of 2022, pushed under a dual GPL/MIT licence.

With the AI buildout also fully underway, Nvidia has crafted the Linux-based DGX OS for its personal mini supercomputing DGX Spark products. This Ubuntu-based system is designed to work with the entire Nvidia software stack – with this AI layer (including software like CUDA, cuDNN and NCCL) being closed.

Although the battle between open and closed source remains active, Nvidia's reversal signalled that the future of enterprise and AI-centric computing was to be found in open source environments.

• Germany advances a Naval laser weapon toward operational deployment by 2029
• Future 100kW laser aims to counter faster and larger aerial threats
• More than 1,000 laser shots validated performance during extended naval trials

Germany has confirmed plans to field a high-energy laser weapon aboard naval vessels by 2029, aimed primarily at intercepting drones.

Defense contractors Rheinmetall and MBDA signed a contract in June 2026 worth several hundred million euros to build the complete system.

The program follows years of testing a demonstrator aboard the frigate Sachsen, which fired more than 1,000 shots during trials.

From demonstrator to deployable weapon

The demonstrator currently operates at roughly 20 kilowatts, sufficient for neutralizing small drones and light surface targets near naval vessels.

Future versions are expected to exceed 100 kilowatts, a substantial increase intended to counter larger and considerably faster aerial threats.

During trials, the demonstrator covered roughly 28,000 nautical miles between the Baltic Sea and the Mediterranean over more than a year.

Another testing phase included over 100 live firing events under maritime conditions, examining tracking accuracy, reaction speed and engagement performance against fast-moving drones.

“The laser weapon system will provide our personnel deployed on naval vessels with a significantly higher level of protection, particularly when it comes to countering drones,” said Roman Koehne, head of Rheinmetall’s weapons and ammunition division.

Beyond countering drones, engineers hope the technology can eventually intercept guided missiles, rockets, and even artillery shells fired from considerable range.

Company officials say the system also demonstrated the ability to strike targets set against open sky rather than solid terrain.

The British, French and Belgian forces are pursuing similar systems — a broader push toward directed-energy weapons across European navies more generally.

The United Kingdom's Royal Navy, for instance, already plans to install its DragonFire laser weapon aboard a destroyer by 2027.

Similarly, Belgium committed €3.1 billion toward layered air defenses including Skyranger 30 systems, GM200 radars and 10 NASAMS batteries.

Costs, sovereignty, and unresolved questions

The German program places strong emphasis on domestic supply chains, with serial production expected largely to occur within Germany.

MBDA Deutschland's managing director, Thomas Gottschild, described the containerized system as an affordable option for guarding ports and other secure facilities.

Neither company has disclosed a final per-unit price, leaving open how affordable the weapon will prove across a full production run.

Scaling the laser from 20 kilowatts to more than 100 kilowatts introduces significant thermal and power-generation challenges aboard confined naval platforms.

However, officials believe naval vessels offer sufficient electrical power, cooling capacity, sensors and space needed for the demanding laser weapon systems.

Both companies describe the system's technology readiness as very high, citing more than a year of continuous shipboard testing under real conditions.

Procurement officials have not specified interim milestones, making it still difficult to assess whether the 2029 goal remains realistic at this point.

Further trials remain necessary before procurement, but recent milestones suggest European navies now view directed-energy weapons as practical complements to existing missiles and guns.

Via Defense News

• Micron's 6600 ION SSD boosts random write performance with an unusually large onboard memory design
• Benchmark testing exceeded several official performance specifications during enterprise evaluation
• Massive 64GB DRAM gives Micron a clear performance advantage over rivals

Micron's new 6600 ION enterprise SSD packs 245.76 TB of QLC flash storage into a single E3.L form-factor drive, and has garnered some high praise in initial reviews.

TweakTown reviewer Jon Coulter awarded the drive a rare 99% score, calling it "the one to beat in its class."

The drive stands out mainly because of its unusually large 64GB of onboard DRAM, an uncommon amount for this capacity class.

Why more onboard memory changes everything

Most ultra-high-capacity SSDs near 256TB, like the DapuStor 245.76TB PCIe Gen5 SSD, use a 16:1 ratio of NAND to DRAM, resulting in only 16GB onboard at this capacity point.

Micron instead uses a 4:1 ratio, giving the 6600 ION a full 64GB of onboard DRAM for indexing random write operations. This larger memory pool lets the drive sustain around 50,000 random write IOPS at queue depth 256, exceeding its 42,000 IOPS spec.

By comparison, competing 245.76TB drives with a 64K IU reportedly manage only about 15,000 IOPS for random writes under similar test conditions.

That performance is reportedly more than three times faster than rival 245.76TB drives built with a smaller 64K indirection unit.

Sequential performance also reaches up to 13,900 MB/s read, and 3,159 MB/s write, both slightly above Micron's factory specification claims.

Random read performance hits roughly 1.78 million IOPS, matching Micron's published specification for the drive exactly under identical test conditions.

These results were measured using an Intel Xeon w7-2495X processor on a PCIe Gen5 platform running Ubuntu Linux, confirming the figures under real enterprise conditions.

Pricing and practical limitations remain unclear

Despite the strong benchmark results, Micron has not published an official price for the 6600 ION at the time of writing.

Reports suggesting a price beyond $100,000 have circulated online, though no listing confirms that specific figure at the time of writing.

Enterprise SSDs at this capacity typically sell through direct vendor contracts rather than public retail listings, making pricing hard to verify.

The drive carries a 5-year limited warranty and supports major operating systems including Linux, Windows Server, and VMware ESXi across enterprise deployments.

It offers a 1-drive-write-per-day endurance rating, a modest figure that still fits typical enterprise storage workloads at this scale.

The drive also includes power-failure protection and full data-path protection, standard features expected in enterprise-grade storage designed for continuous operation.

The benchmark numbers alone do not confirm real-world value, since pricing and support costs remain unknown.

Whether the 6600 ION genuinely leads its category depends heavily on how competitors price similar high-capacity QLC drives over the coming months.

Until official pricing becomes available, claims about its market position should be treated as preliminary rather than fully confirmed.

RayNeo X3 Pro: 30-second review

RayNeo, the AR glasses arm of TCL, launched the X3 Pro globally in December 2025, following a well-received debut in the Chinese market. It represents the company's most ambitious product to date: a standalone pair of AI-powered augmented reality smart glasses that aims to put a useful, persistent digital layer over your view of the world, without requiring you to carry a tethered compute unit.

The headline hardware is the dual-eye full-colour MicroLED display, powered by RayNeo's own 'Firefly Optical Engine' and delivered through waveguides co-developed with Applied Materials. With 6,000 nits of peak brightness and 16.77 million colours, it is probably the best display currently available in any smart glass product, eclipsing even the Meta Ray-Ban Display's 5,000-nit panel. The simulated image is equivalent to a 43-inch screen viewed from two metres, within a 30-degree field of view.

Under the frame sits a Qualcomm Snapdragon AR1 Gen 1 processor — the purpose-built platform for this class of device — paired with 4GB of RAM and 32GB of onboard storage. The X3 Pro runs RayNeo's AIOS, an Android-based operating system, and is integrated with Google Gemini 2.5 (Beta) for multimodal AI assistance. A 12MP Sony IMX681 sensor handles photography and 4K video, accompanied by a secondary monochrome camera for spatial positioning and depth tracking with 6DoF + SLAM support.

At 76 grams, the X3 Pro is lighter than the Inmo Air 3 (119g), and only a few grams heavier than the Meta Ray-Ban Display (69g). The frame is built from an aerospace-grade magnesium-aluminium alloy, and control is handled via a five-way touch panel on the right temple, with support for Apple Watch gesture control promised in a future OTA update.

The device's single greatest limitation is its 245mAh battery. Under light use, you may approach three to five hours. Under active use that might be navigation, AI queries, camera recording, or app usage, the running time plummets to as little as one or two hours, and it can be as little as 45 minutes. The only saving grace is a recharge of around 45 minutes via USB-C.

At $1,169, the X3 Pro is a premium early-adopter product with genuine technological credibility, but a hefty price tag. The display alone makes a compelling case for the future of AR glasses. Whether that future is worth over a thousand pounds to experience today is a question each buyer must answer for themselves.

RayNeo X3 Pro: Price & availability

The RayNeo X3 Pro launched globally in December 2025, initially priced at $1,099 on an early-bird basis, rising to $1,299 at standard retail.

At the time of writing, RayNeo sells direct from its website here, with delivery to the United States, the United Kingdom, France, Italy, Germany, and other markets.

In the UK, the full retail price is £1,169, and in the USA it’s $1,169. Considering that the exchange rate on the day of writing is $1.34 to the pound, UK customers pay roughly 25% more for the same products for no obvious good reason.

Prescription lens inserts are available separately from around $49 / £49, supplied through RayNeo's partner Lensology.

What’s a little odd is that these glasses aren’t available on Amazon.com, when almost everything else RayNeo makes is.

By way of comparison, the Meta Ray-Ban Display starts at $799, the Even Realities G2 at $599, and the Halliday Smart Glasses at $500. Traditional smart glasses without a display, such as the Ray-Ban Meta, are available for considerably less.

The X3 Pro commands a significant premium, but the technical specification that includes the dual-eye MicroLED display and the Snapdragon AR1 platform is a major step up from those alternatives.

RayNeo also offers existing X-series customers a 'RayNeo Explorer' lifetime benefit, providing a $200 discount towards future X Series purchases.

What colours my perspective on the price is that these aren’t dual-purpose glasses that can also be used to watch movies. They’re only for AR, which makes the high price even harder to justify.

• Score: 3/5

RayNeo X3 Pro: Specifications

RayNeo X3 Pro: Design & build

• Divisive aesthetics
• Wearability issues

There is an obvious problem with products like the X3 Pro, which is that the design telegraphs that these aren’t just glasses, drawing attention to the wearer.

This tension between engineering achievement and social wearability is perhaps the defining characteristic of this first generation of capable AR glasses, and the X3 Pro didn’t dodge that bullet.

The frame takes broadly Wayfarer-style cues, i.e. being thick, squarish and dark. In short, these look like John-Paul Belmondo wore them at the end of the 1960s, before Michael Cain borrowed them to play the classic British spy, Harry Palmer.

That might be delightfully retro, but two cameras sit in the bridge between the lenses, a small indicator light sits on the front frame (active when recording), and distinctive protrusions near the temple hinges house the MicroLED projectors, giving the game away.

The temples are noticeably thicker than conventional eyewear, accommodating the speakers, electronics, and battery. The USB-C charging port sits at the tip of the right temple.

Structurally, the X3 Pro is more refined than its predecessor, the X2 Pro. RayNeo cites eleven structural optimisations and the use of aerospace-grade magnesium-aluminium alloy to achieve a 36% weight reduction over that earlier model. The result is a frame that, at 76g, sits comfortably on most faces without the ear pressure or nose strain that plagued heavier competitors. Multiple reviewers noted that they occasionally forgot they were wearing them during prolonged use.

The lenses themselves have good optical transparency when the display is off, meaning the world doesn't take on the tinted quality of sunglasses during non-display use. Interchangeable nose pads in different sizes are included, and prescription lens inserts are available through Lensology.

For older people, me included, these lens inserts are a necessity, since the glasses require abnormal eye-muscle acrobatics that those without perfect vision are unlikely to achieve without some help.

Fit adjustment is largely limited to nose pad selection, which does tend to put more pressure on the bridge of the nose. Previously, with the Air 3s Pro, RayNeo offered adjustable temple angles, but these aren’t available on the X3 Pro.

And, because of this, depending on your face shape, you can find that the display is dramatically offset from the ideal line of sight. As I’ll talk about later, I had big issues with this, and it made using them extremely difficult.

In short, if social discretion is a priority, this is not the device for you. If you are the sort of person who wears technology proudly, or who has a professional or specialist use case, the design is functional as long as your face and eye geometry fall within a specific envelope.

• Design & build: 3.5/5

RayNeo X3 Pro: Features

• Impressive display technology
• Sony IMX681 camera sensor
• Tiny battery

The X3 Pro's MicroLED dual-eye display is, by wide consensus, the standout feature of this device. Unlike single-eye displays used by some competitors, the X3 Pro projects identical imagery to both eyes, producing a more natural and immersive AR experience that doesn’t assume binocular compensation on the viewer's part. The 640 × 480 resolution per eye is modest by smartphone standards, but it is appropriate for a heads-up overlay and is rendered with genuine clarity at the 30-degree field of view.

Peak brightness of 6,000 nits is a notch above the Meta Ray-Ban’s 5,000 nits, making the display legible in direct sunlight and suitable for navigation or outdoor use. These aren’t meant for media consumption, and therefore don’t include shields to reduce external views, so the graphics need to be bright.

The display sits centrally in the wearer's field of view, rather than in the lower-right corner (as on the Meta Ray-Ban Display). This means AR content is more prominent and easier to read, but also more obtrusive. You cannot easily consume AR content passively while doing something else. It is a deliberate design choice that suits dedicated, task-focused use over an ambient, always-on overlay.

The primary camera uses a Sony IMX681 sensor capable of 12MP stills and 4K/3K video. A secondary monochrome camera assists with spatial positioning, depth tracking, and dual recording. In daylight conditions, camera output is described as decent, with the wide-angle field of view well-suited to point-of-view recording.

But in low light, there is a tendency to visible noise and graininess, and the lack of digital zoom or manual camera controls reduces flexibility. A recording indicator light on the front frame activates when the camera is in use, serving both as a privacy indicator and a practical reminder.

These don’t take pictures that would worry any mid-tier phone, and most entry-level Android phones have better sensors.

The X3 Pro's battery life is probably its greatest limitation, since a 245mAh cell is simply not large enough to support extended active use of the device's headline features.

RayNeo's claim of up to five hours applies to very light use, and by that, they probably mean music playback and limited screen time. In practice, active use scenarios significantly reduce this figure. In a few of my sessions, the time was a fraction of that amount, and the worst offenders for eating battery capacity were translation, video capture and navigation.

In theory, you could have a hip-mounted power pack attached to the USB port of the X3 Pro, but when I tried this, it pulled them out of square and made reading the display even harder.

Thankfully, the glasses do feature wear detection, automatically powering down when removed. This helps conserve battery during breaks, but carrying a power pack around is practically a necessity if you intend to use them for any extended time.

Contextually, the limited battery is an inevitable consequence of the 76g weight target. A larger cell would mean a heavier device. RayNeo engineers have made a considered trade-off here, and future hardware iterations will presumably seek to improve energy density. It may be that the makers can engineer better power management through firmware adjustments, but with only 245mAh of battery to work with, there is only so much that can be done.

Without a doubt, the primary reason to hesitate before purchase is battery life.

• Features: 3.5/5

RayNeo X3 Pro: Software

• Google Gemini 2.5 (Beta)
• Live translation
• Side-loading apps

RayNeo AIOS, the operating system built on Android and structured around four primary screens: a home screen showing time and status indicators, a quick-actions panel, an app launcher, and a notifications screen. Navigation is via the five-way touch panel on the right temple, with voice commands available via 'Hey RayNeo'. The interface is responsive and, for the constrained form factor, relatively intuitive.

According to RayNeo, it's Google Gemini that’s the flavour of AI baked into AIOS, and I suspect it’s Google Gemini 2.5 (Beta), which is a long way behind the current models that Google is promoting.

Compared to some other talking AI’s I’ve used, this one is pretty average. For starters, even though I’m in the UK, it insisted on using a chirpy American accent. And, if I asked what the temperature was, the answer arrived in Fahrenheit, times in a 12-hour clock and distances in feet and inches. Yes, Gemini, the world is America.

But aside from being fixated on a region that’s more than 3,000 miles away, the other issue was that it got simple questions wrong from the outset. As it loves America, I asked it to name the last ten U.S. leaders. It got the name and the order correct and then fumbled the answer by saying that all these people had been President in the past ten years.

I tried to subtly nudge it in the right direction by asking which ones were the President in the past ten years, but it failed to notice the discrepancy between what it was saying now and what it said previously.

Thankfully, it didn’t fall for the classic "walk or drive" question for the car wash, but I think all AI platforms are hardwired to answer that way, since it’s an obvious pitfall.

Compared to the latest versions of the major AI providers, the AI on this platform isn’t going to write Skynet anytime soon.

Real-time translation is a more advanced feature, supporting 14 languages and delivering approximately 2.1-second response times. Translation can be delivered as on-screen text or synchronised audio. In testing by other reviewers, accuracy was broadly good, though the system waits for the speaker to finish before translating. That’s necessary in some languages, like German, but it does come across as a less-than-natural conversation and can feel stilted.

Navigation is powered by HERE WeGo Maps (used by BMW and Audi), projecting turn-by-turn directions and nearby landmarks directly into your field of view. This is one of the most practically compelling use cases for the device, eliminating the need to look down at a phone while on foot. Unfortunately, the app never loaded on my glasses. Every time I tried to download and install it, it failed. Other apps were installed, so I’m unsure why this one refused to.

I know that some software for this device requires side-loading, which isn’t something many users will be happy to perform.

• Software: 3.5/5

RayNeo X3 Pro: Performance

The Snapdragon AR1 Gen 1 is purpose-designed for augmented reality applications, and the X3 Pro benefits accordingly. Day-to-day navigation, AI queries, notification handling, and app use are smooth under normal conditions. The combination of 4GB LPDDR5 RAM and 32GB storage is appropriate for the use cases the device targets.

That said, compared with a modern smartphone, this isn’t the most powerful platform, and with some more resource-intensive tasks, the cracks start to show.

The glasses do support 6DoF + SLAM with Falcon Image spatial positioning, and the AR overlay alignment is typically accurate and stable under testing. But the issue here is more about how close this platform is to being overrun, and there are hints it's not ever too far from the edge.

But this reviewer had many more issues with this device, which is partly why I waited more than six months before completing my review.

When I first got these in 2025, they did almost nothing. Since then, the firmware updates and enhancements that come via the mobile app have transformed the functionality provided, but they haven’t addressed some of the issues I’ve had from the outset.

The first big problem I had was seeing the projected images, not because the glasses didn’t work, but because they were almost out of my field of view. Some of this was my long-sightedness that made the images seem soft, but I couldn’t see the entirety of the display without balancing the glasses on the very tip of my nose. If I didn’t do that, the image would have been presented as below me and barely in sight. Lifting the glasses to make the image central causes it to disappear.

I’m not confident that spending another £50 on the proper lenses would fix that issue.

That made just seeing things a challenge, but the other issue I had was using the touch panels on the sides of the glasses for directing the interface, because half the time they just ignored my instructions or did something I didn’t ask for. In one instance, I deleted the To-do application from the glasses, not because I wanted to, but because the glasses took one swipe as my instruction to do that, and then refused to cancel that erroneous request.

I did consider getting a small hand controller to make it easier to use or even using the phone as a touchpad, but frankly, at this price, it should be easier than it was.

My final complaint about this device is how some aspects aren’t thought through. One of the apps is a translation app, and you can stand in front of a person from another country and get real-time translation of what they are saying. And, it works. You can even run a YouTube video of someone speaking another language and see it translated.

However, my problem is how you might use this in the context of being a tourist in a foreign country. Let’s imagine I’m in Japan, where they speak a language I don’t, and I walk into a shop where a sales assistant asks, ‘What are you looking for?’ I understand this, because the glasses translate for me, but it can’t reply in Japanese.

At which point, phone translation, where you can show the person your reply in their language, or it speaks for you, works much better. Obviously, if you like to sit on a train and listen to people gossiping about you in their language, thinking you can’t understand them, it's great, but it seems an expensive device to do just that.

• Performance: 3.5/5

Should you buy the RayNeo X3 Pro?

The RayNeo X3 Pro is, technically, the most impressive pair of smart glasses currently available for purchase. The dual-eye MicroLED display is genuinely impressive, with bright enough images for outdoor use, colourful, and binocular in a way that no other glasses at this price point can match.

The integration of Gemini AI would make it genuinely useful beyond being a novelty if the model were newer and didn’t assume that all English speakers are Americans.

The camera produces capable results in good light, and the 76g weight is a remarkable achievement for the hardware it contains.

But the £1,000+ price tag demands honest scrutiny of what you're buying, and the answer is: a first-generation product. The battery will frustrate most users who intend to use its headline features for more than an hour or two at a time. The app ecosystem requires technical workarounds.

The aesthetic is also overly conspicuous, and considering how people are quite rightly objecting to unwanted image capture and AI in general, expect some push-back from others if you wear these in public.

For early adopters, AR developers, and professionals with specific use cases, such as live translation, heads-up navigation, and meeting transcription, the X3 Pro is credible but far from perfect. For mainstream buyers hoping for an all-day, all-purpose wearable, the technology is not quite there yet, but this is the clearest indication yet of where it is heading.

Buy it if...

Don't buy it if...

If you want genuine 8-core desktop power in a footprint you can tuck behind a monitor, this Bosgame P6 mini PC for $520 (was $640) at Amazon is worth checking out, with the machine built around AMD's high-performance Ryzen 9 6900HX.

The Ryzen 9 6900HX is the real draw here. It's an 8-core, 16-thread chip built on a 6nm Zen 3+ process, with a boost clock up to 4.9GHz. That puts it well ahead of the Ryzen 5 and Ryzen 7 mobile chips that populate most mini PCs in this price range.

The Bosgame P6 is genuinely capable of sustained multi-threaded work like video encoding, compiling, or running several lightweight VMs at once without slowing down.

Today's top mini PC deal

The Radeon 680M integrated GPU pairs with that CPU well. It's an RDNA 2 part with 12 compute units, and while it's not a substitute for a discrete card, it comfortably handles 4K media playback and esports titles like CS2 at high settings — useful if this doubles as a casual gaming box alongside office duties.

24GB of LPDDR5X-4800 memory is a healthy amount for multitasking, though it's worth noting this RAM is soldered to the board and not user-upgradeable — what you buy is what you're stuck with. The 1TB PCIe 4.0 SSD fares better: BOSGAME includes a second M.2 slot, so storage can be expanded up to 4TB down the line.

Connectivity is a genuine strength at this price. Dual Gigabit Ethernet ports let you keep two networks separated or add load balancing — handy for home labs or NAS setups — while Wi-Fi 6E and Bluetooth 5.3 cover the wireless side. Triple 4K display output via HDMI, DisplayPort, and a full-function USB-C port (with data, Power Delivery, and DisplayPort Alt Mode) rounds out a well-connected little machine.

A couple of honest caveats: the soldered RAM means there's no upgrade path if 24GB becomes tight later, and Bosgame's mini PCs typically ship with an external wall-socket power brick rather than a laptop-style supply, so factor that into your desk setup. If you need discrete-GPU-level graphics performance for gaming or GPU-accelerated creative work, this integrated setup won't get you there either.

For an 8-core desktop replacement with strong connectivity, the BOSGAME P6 is a solid pick in this category.

Want to do more with your machine? We tested the best mini PCs and selected the top mini PC deals for video editing and gaming that you can get right now.

More mini PC deals

• Shop other mini PC deals at Amazon.

• Meta's 4,000 acre Hyperion data center will receive an additional $40 billion investment
• Meta says the project has been a huge benefit for the local community of Richland Parish
• Multiple groups and local residents have criticized the massive campus and its energy sourcing

Meta has announced that an additional $40 billion will be used to fund the construction of its enormous 4,000 acre data center campus in Richland Parish, Louisiana.

This additional funding brings the total Meta investment for the ‘Hyperion’ site to over $250 billion when combined with Bloomberg reporting of $200 billion allocated for the project in May, plus a previously announced $10 billion investment for the data center and surrounding community.

The data center itself is expected to consume 5 gigawatts of computing power, with an additional 2 gigawatts for wider campus needs, placing the predicted total usage upon completion at more than three times the power consumed by the city of New Orleans.

A boon or bane for the local community?

In a Meta blog post, the company boasted about the benefits of the enormous data center campus for the local community.

It cites $50,000 bonuses for local teachers thanks to increased tax revenues on the Hyperion campus, alongside $1.6 billion in contracts for local businesses and other funding from Meta for public schools and youth programs via the Data Center Community Action Grants project.

“It’s life-altering for our teachers and their families, and it’s transforming our schools. Meta’s investment has made Richland Parish a destination for education as well as industry,” Richland Parish’s School District Superintendent, Sheldon Jones, told Meta.

Meta also said that once the project was completed, it would support 1,000 jobs. There is also a further $1 billion earmarked for investment in local infrastructure improvements, including roads, water and wastewater systems.

A promotional video accompanying Meta’s local funding news shows the story of a backwater town faced by a lack of business and opportunities, with local residents, teachers, and business owners speaking of the positives of having the Hyperion campus being constructed, accompanied by upbeat orchestral music.

While the select individuals featured in the Meta PR video preach of the benefits, there has also been local opposition to the project, which has been steeped in controversy.

The other side of the coin

Multiple environmental and consumer groups have pushed back against the project since its announcement, particularly around how the power for the data center will be generated. Entergy Louisiana is spending billions to construct 10 new gas-fired power plants to provide the necessary electricity. But many communities across the US who live in the vicinity of such power plants have reported a wave of symptoms, such as dizziness, nausea, vertigo, and sleep disruption.

A Floodlight investigation, published by the Guardian, found that the success of Hyperion’s approval rested largely on the shoulders of a single Louisiana state senator, John “Jay” Morris.

Many of the land purchases and sales required for Entergy to construct the planned gas turbine plant were undertaken by Morris and his partners, Floodlight claims. Louisiana law requires government officials to recuse themselves from voting when a conflict exists, and prohibits the use of public office for private gain.

Morris has denied any wrongdoing. “It makes a nice story if you can try to show that I have some sort of conflict. But under Louisiana’s ethics laws, I don’t,” Morris told the Guardian.

The Union of Concerned Scientists (UCS) and the Alliance for Affordable Energy (AAE) have both criticized the Hyperion site, especially the amount of power it is planned to consume, warning that the site will unfairly shift the cost of electricity and infrastructure upgrades on to the site.

Additionally, the approval for the Entergy gas turbine power plant was fast-tracked through the Louisiana Public Service Commission, with the UCS warning that Entergy ratepayers would have to cover the costs of a $550 million transmission line that is only necessary because of the Hyperion site, alongside the fuel costs for the new gas turbine plant.

The Bambu Lab A2 L is a printer I’ve been expecting for some time. Essentially, it’s a larger A1, just with technology tweaks and, of course, scale. The big feature here is a 330 x 320 x 325 build area, making it one of the largest in its class of large open 3D printers.

While the A1 is a superb printer for beginners, there’s often a feeling that the print bed size can be a little bit restrictive, especially for some enthusiasts as well as businesses looking to model, create costume items or larger-scale prototypes, and this is what the A2 L’s larger print bed essentially facilitates.

What I like here is that Bambu has essentially created a large-frame design that keeps costs down compared with the more expensive, professional-level H2, P2, and X2 designs. This is more of the entry-level, but still it offers all of that technology and refinement that Bambu Lab always offers, and whilst many people will look at the open-frame design as a little bit of a negative point because you won’t be able to use engineering filaments, if the bulk of what you do is basic print farm work or you’re just printing PLA or basic TPU, then this machine is going to be a perfect option.

During the test, another point that came up was that the Wi-Fi is limited to 2.4 GHz. Since I was testing other kit at the time and didn’t have that 2.4GHz bandwidth, I couldn’t quite figure out what was happening. Only when I went for the dual 2.4 and 5GHz did it connect okay, but really, this is a minor thing for most businesses, and that would be absolutely fine.

The build area is obviously the major difference between this and the A1: the A1 has a nice square 256 x 256 x 256 build area, whereas this one upgrades to 330 x 320 x 325. I do kind of wish that Bambu Lab had kept an absolute cube build, as it just makes things easy when you’re orienting; however, it’s nice to have such a large space to work with.

Along with a larger build volume comes a new extrusion system, fully upgraded from the A1 standard direct-drive setup. On the A2, it adds a PMSM closed-loop servo extrusion system, designed to deliver smoother extrusion and monitor for issues during the print process. Another good upgrade is the motion stability system, which is far more advanced, as you’d expect given the couple of years between releases. Part of the reason for this is not just the passing of time, but also the fact that, because this is a much larger machine, it’s more prone to errors, so this part is finely tuned.

Further aiding error-free printing is a new package: physical blob and clogging detection, PMSM extrusion monitoring, and all live alongside all the safeguards the A1 featured when it was released.

One of the other features I have been really keen to see is the creative tool expansion, which essentially means, like the H2D, you can pop on an optional blade-cutting or pen-plotting module, so you can also work with stickers, card, leather, fabrics and a whole range of drawing and plotting projects alongside the 3D printing.

Really, this is a machine designed to sit there and get on with the work out of the box, and for less than £500/$500, you have a machine that can print multi-filament easily. If you do want to print engineering materials, you'll just have to spend a little bit more on one of the enclosed machines, but for most people, given the price, this is actually a fair option.

Bambu Lab A2L: Price and availability

The Bambu Lab A2L is available in the US for $469 direct from Bambu Lab - with a Combo version priced at $569.

It's also available in the UK for £319 at Bambu Lab, with the Combo version costing £429.

Bambu Lab A2L: Design

It’s obvious that the A2L is in the same family as the A1. Essentially, it looks similar, just on a larger desktop-style frame, that measures 544 x 529 x 505mm, quite a substantial footprint, and it’s no lightweight at 12.8kg.

That scale of machine offers the current build area of 330 x 320 x 325mm, which is substantial, and in order to ensure proper rigidity for this large-format printer, the construction is made from aluminium with steel for the chassis, and a plastic outer frame to make it look nice rather than the look of printers past. To be honest, with Bambu Lab's usual design flair, this really does look like a great machine.

One thing to note is that this is an open-frame construction, and there is not, and probably will never be, an enclosure for it. There are higher-level machines like the Super P2S if you do need to utilise more engineering materials, and they don't cost a lot more.

What really appeals to me about this design is the optional cutting-and-drawing module that can be attached. This essentially boosts the machine's usability, and if you’re into crafting or modelling, as well as working in a small design business, having all these optional extras in one machine makes a lot of sense. One point to note here is that, unlike the H2D, which offers both of these modules to expand its functionality, it is slightly stripped down here, and for safety reasons, there’s obviously no laser module.

While it’s pitched to a wide range of users, it's primarily for a home enthusiast, although I’ve already spoken to several people who run print farms and are excited about this machine because of the sheer volume of prints they’d be able to produce on each print platform. Having run successive tests over a one-month period, I can confirm that the reliability for this type of use makes this a really great option, even in the commercial business sector, as long as you don’t want to print with engineering materials that require an enclosure.

Like all Bambu printers, it's designed for multi-filament printing and is fully compatible with both AMS and AMS Lite units. If you do purchase the combo unit, then you’ll get the AMS Lite, as you did with the A1. With this machine, you can link up four AMS units to get up to 19 colours.

Bambu Lab A2L: Features

The first thing that should really be said about the A2L is that it’s not a huge leap in technology, although there are improvements across the board, mainly to facilitate its larger build area, more than material support or additional new features. So, firstly, the headline is the scale of this machine, and as stated before, the 330 x 320 x 325mm build volume is ideal for large-format printing. If you’re thinking about cosplay or prototyping larger objects, then, to be honest, for the price, there is really very little at the moment that can match this.

That larger size usually means it’s much more prone to errors, and whilst the larger H2 machines are in line with the print volume this machine offers, the one big difference here is that it is open-frame. That not only makes it much cheaper but also restricts the types of materials you can utilise. Essentially, day-to-day prototyping in PLA or one of the other more common materials is absolutely fine, and you can pop in a hardened nozzle if you do want to print with something like PLA-CF or PETG-CF. However, if you want to print with engineering materials, such as nylon, then you are going to have to look at an enclosed machine.

There are several features that really stand out that will help make the printer as reliable as possible. The first is the aluminium-and-steel construction of the frame and chassis, which results in an ultra-rigid unit. During the test, I was surprised to see that, even when printing flat out, it appeared just how stable it was. It is, of course, always a good idea to have a really good, solid table.

Built into the extruder is one of the major upgrades over the A1: a PMSM closed-loop servo extrusion system that offers smoother filament extrusion than the direct drive we saw on the A1. There’s also adaptive vibration compensation with multiple calibration and load-adaptation options, again helping with filament flow. Another surprising thing is that they’ve added granular dampers to the frame. This is essentially to cope with the high speed that this printer is capable of, and it’s worth noting that, at full pelt, this machine is quite ferocious in the corner of the workshop.

As with the A1, blob detection, run-out detection, clog detection and tangled-spool detection are all pretty much standard features.

Some of the standout features are the optional blade-cutting and pen-plotting modules, which expand the machine's capabilities and are ideal for crafters and enthusiasts, especially at this price point.

Once again, there’s full support for the AMS Lite system, and if you buy four AMS units, then you can print up to 19 colours, which is an impressive amount for a base machine at this price.

Bambu Lab A2L: Performance

Getting started with the Bambu Lab A2 L is relatively quick and straightforward. Unlike the fully enclosed machines that we’re starting to get used to from the company, this does require a little bit of construction, but ultimately it’s just the vertical frame with the print head pre-installed, and once you’ve bolted it down and connected a couple of cables, you can then run through the initial calibration process.

This used to be the faffy part, and construction took some time. Now, the construction takes only 5 to 10 minutes; however, once you switch the machine on, you need to connect to your Wi-Fi before it can download and install any updates. You then have to run through the calibration process, which is fully automatic, so just let it run its course; it can take almost an hour on the first run. After that, it should be pretty much self-sufficient, with an automatic reminder alerting you as to when it needs recalibration or when the rails need a little oil.

As I ran through the first prints, all stored on the small MicroSD card, I was impressed by the machine's speed and robustness. Whilst it’s just an enlarged A1 with refinement, the rigidity feels just as good as that of the smaller machine, and Bambu Lab has really made quite a lot of adjustments to the base design in order to facilitate ultra-strong joints that don’t create any wobble or movement through the print process, which is definitely needed at maximum speed.

As the first few prints rolled off the print bed, they had that unique Bambu Lab look, with very fine layering and superb overall surface and inner print quality. Single-filament prints were exceptionally quick, and I checked the models as they were produced. I could see that the fine flow control was exceptionally good, with fine detail printed well.

The point I was going to see, especially with the test models, was that the overhangs and bridges seemed well handled, demonstrating that the new head is capable of producing enough cooling air to set the filament quickly throughout the print.

After checking a series of prints, from those just for fun to more functional parts, in both PLA and PETG, I was really impressed with the quality. Realistically, if you put the prints from this machine up against those from the H2D, which is obviously far more expensive, it’s actually extremely difficult to tell the difference. Whilst this machine is cheaper, the print quality on a purely visual side-by-side basis is identical. Realistically, at the base level, you are just paying for that enclosed space, but then it is obviously a lot more than that. However, for most people, that’s really all they need: a good, large machine for printing in these base materials, which, to be honest, is what most of us use.

If you’re working in design and you need to print out a large prototype, as I have with an air cooler this summer, then this is a really great option. For the models I was printing, it printed at full height, so I could quickly prototype an entire product without breaking it down into too many parts. If you do need to break it down into parts, there’s enough out there to support plenty of different models at the same time.

During the print process and testing, I printed solely for quality purposes, then looked at printing on sale to see how many models I could reliably print, and moved on to more functional models, including an air-cooling system and the associated parts. I am also using the printers to create many props and costume parts for various projects, and, again, the additional size offered by the Bambu Lab A2 L is extremely useful.

What I like about this printer is that, it’s large, and it printed large reliably. I have three or four large-format printers, all of the open design from various manufacturers. However, most regularly fail, whereas the A2 L was consistent with those larger prints, time after time. It’s also large enough for a fully wearable helmet, which is now my go-to test when testing print volume.

A quick check of the 3D test model, and to be honest, it’s as good as any. This test, which used to challenge so many printers, really now just highlights one or two printers that aren’t quite so good, and we just see a little bit of surface imperfection, which means that you need to have a little bit more oil on one of the rails, or give the machine a good clean or calibration, rather than anything more fundamental. Re-printing a 3D Benchy, it came out absolutely perfect.

Moving on to the more in-depth Kickstarter Autodesk 3D printer check, the model produced again looked good. I was a little surprised that there was only a bit of stringing on one of the overhangs, but otherwise the model itself looked exceptionally good, with fine flow control and dimensional accuracy that really highlighted what it was capable of. Whilst this is the larger of the two relatively entry-level models from Bambu, it’s still very, very capable.

Further, to put this to the test, I ran it for two weeks nonstop, producing multiple parts of the same model to check reliability. What I was interested in seeing was that after a certain number of hours, the interface asked me to maintain the rails. This is just a good reminder that Bambu Lab is really looking after the machines and notifying users when maintenance is required. I was also pleased to see that printing just pure PLA from four reels was consistent. I am not a huge fan of the AMS Lite four-reel design; however, there’s no doubting that it works. Here again in the workshop, it produced prints without issue.

The Bambu Lab A2 L is a natural progression as a 3D printer. It’s not one to get overly excited about; it is what it is and will enable you to print large. It’s relatively inexpensive. Budget-wise, what impressed me was just how reliable it proved to be, and with that large build area, you can not only print more but also print at a higher volume, at one of the lowest prices for this level of reliability and quality on the market.

If you run a print farm or you prototype goods, or you’re just looking for your next larger-format printer for producing cosplay costumes, then I can see the A2L being a popular choice that gives you far greater flexibility than its smaller and now slightly older sibling.

Bambu Lab A2L: Print quality

Dimensional accuracy - score of 4/5

Target 25 = X: 4.88mm / 0.12mm Error | Y: 4.82mm / 0.18mm Error
Target 20 = X: 9.55mm / 0.45mm Error | Y: 9.89mm / 0.11mm Error
Target 15 = X: 14.86mm / 0.14mm Error | Y: 14.83mm / 0.17mm Error
Target 10 = X: 18.84mm / 0.16mm Error | Y: 19.81mm / 0.19mm Error
Target 5 = X: 24.81mm / 0.19mm Error | Y: 24.71mm / 0.29mm Error

X Error Average = 0.212
Y Error Average = 0.188
X&Y Error Average = 0.2

Fine Flow Control - score of 5
Fine Negative Features - score of 5
Overhangs - score of 5
Bridging - score of 4
XY resonance - score of 2.5
Z-axis alignment - score of 2.5

Adding up the totals gives a final score of 28 out of 30

BambuLab A2L: Final verdict

When you are looking at the A2L, you have to think about the build quality, given its size, and what I’m really pleased to see is, firstly, that Bambu Lab has continued with its design principles. This is a large-format printer that looks great. It has all the features from Bambu that many have come to expect.

Set-up and maintenance are exceptionally easy. As long as you follow the prompts, you’re going to have trouble-free printing. The only thing I would say, like with Bambu Lab’s other printers, is that it can be a little overprotective of the print: at any sign of string or debris on the build plate, it will stop, and you will have to check it. It won’t just plough through, which in itself is actually a good thing, but I can’t help but feel it’s just a little bit oversensitive.

This is still an open-frame design, so whilst the size will definitely appeal to professionals and high-end enthusiasts, you do need to consider that it will only print materials such as PLA, PETG and specialist TPU. Anything you would probably choose to make functional parts does require an enclosure, so from that point of view, it’s a little limited. However, that’s not what the A2 series is about, and actually, as a large-format printer for printing PLA prototypes, costumes and all manner of other objects in that fashion, it’s a great choice.

The other point about this 3D printer is that it has printing and cutting ability with the pen and cutter, which lends it to even greater use. Whilst this is a good option and a space-saver for many, I definitely prefer a cutter that draws the material through rather than one that pushes it down, as this machine does. Still, for my stuff, as long as the blade is exceptionally sharp, it works very well for both cutting and plotting.

If you own a company and want to create large-scale models or PLA, it’s a brilliant 3D printer. In my tests, I found it to be good and reliable, except that it stops on almost anything. But a quick check, either physically or through the Bambu Handy app, and get it set going. If you’re a keen model crafter or enthusiast of any type, this is a superb option, as long as you’ve got that slightly larger space to keep it. If you’re working in product design and you just need a large-scale printer for prototyping, this is again a superb choice.

Should I buy the Bambu Lab A2L?

Buy it if...

Don't buy it if...

For more models, I've tested the best 3D printers you can get right now.

• Stripe OLT found ModHeader v7.0.18 carried a hidden spyware SDK, exfiltrating visited domains daily to a Chinese‑owned server and acting as adware
• The extension had 1.6M downloads across Chrome and Edge before being pulled but installed endpoints remain at risk
• Researchers urge defenders to identify and remove existing installations, as removal from stores does not automatically remediate compromised devices

ModHeader, a trusted Chrome and Edge browser extension with more than 1.6 million downloads, was found to be malicious, apparently sending sensitive data to a Chinese-owned server, and has since been pulled on both repositories.

Security researchers Stripe OLT revealed the news in a new report, outlining how a ModHeader build v7.0.18 carried a hidden spyware SDK.

As per Stripe OLT, the spyware collects domains users visit, encrypts the data with AES-GCP, and then sends it - once a day - to a remote server. The collector was found inactive by default, but the required code, encryption key, and upload schedule were all already embedded in the extension.

Links to Chinese actors

Researchers found no command-and-control functionality, which means the server only receives the stolen data and cannot communicate back. The extension also worked as an adware, displaying ads and opening advertising tabs on updates, including on enterprise-managed devices.

The researchers attributed the attack, albeit with low confidence, to a Chinese-speaking threat actor. The exfiltration domain routes emails through Lark, which is a suite common with Chinese-speaking teams, it was said. They also found Chinese strings in the code, and said that the listing ships a Simplified Chinese locale.

ModHeader is a Chrome and Edge browser extension that allows users to modify HTTP request and response headers sent between their browser and websites. Developers and security researchers use it to test APIs, troubleshoot applications, and simulate different environments. It has around 900,000 users on Chrome, and another 700,000 on Edge.

According to The Hacker News, Microsoft pulled the tool from its repository on June 3 2026, followed by Google a week later, on July 10.

“Following our disclosure, Google has removed the extension from the Chrome Web Store,” Stripe OLT concluded. “We welcome this action, but removal from the store does not automatically remediate endpoints where the extension was already installed, so defenders should continue to identify and remove existing installations.”

Somewhere between "I need a new mouse" and "I have recreated a 1985 living room on my desk," there's a line, and 8BitDo's Retro lineup will happily walk you right over it.

The company's NES-styled mouse, keyboard, and numpad are built on the same off-white-and-red aesthetic, they pair over the same software, and stacked together, they turn a boring desk into something that looks like it should be plugged into a CRT.

While they're sitting at full price - and from what I've seen, rarely get a discount - the prices are still refreshingly reasonable for what you get. For the ultimate retro home office setup, this is an easy recommendation from me.

Go Retro

Why this combination works

The trick with 8BitDo's Retro line is that the nostalgia isn't hiding a corner-cut product underneath it. The keyboard uses genuinely hot-swappable Kailh switches and PBT keycaps, which is the kind of spec sheet you'd expect from a keyboard enthusiast brand, not a novelty NES tribute. The R8 mouse's PAW 3395 sensor and 8,000Hz wired polling are legitimately competitive with mice that don't have a charging dock shaped like a museum display stand. And the numpad, which could easily have been an afterthought, gets its own little identity with the calculator mode.

Buying all three together also solves the problem that usually comes with retro-themed gear: matching. A single retro-styled keyboard on an otherwise modern desk can look like a costume prop. A full matching set — same grey, same red accents, same rounded plastic — reads more like a deliberate aesthetic choice than a novelty impulse buy, and the shared 8BitDo Ultimate Software means the mouse and keyboard's macros and profiles live in the same app rather than three different pieces of bloatware.

An honest take before you commit to the full set: the keyboard's Kailh Box White V2 switches are genuinely loud — reviewers have clocked it noticeably louder than a typical mechanical board, so it's not the pick for a shared office or a thin-walled apartment unless you plan to hot-swap in quieter switches down the line. The numpad also doesn't yet support macro programming through 8BitDo's software, so for now it's a numpad and calculator first, macro pad second. And obviously, the whole appeal here is the aesthetic — if the retro look isn't your thing, the money is better spent on gear with a more neutral design.

For anyone who wants their home office to look like it moonlights as a 1985 living room, this trio is about as complete an NES-themed desk kit as you can currently assemble.

• See more from the lineup at 8BitDo's Amazon storefront.

• ShinyHunters abused OAuth trust in Salesforce by tricking users and later compromising SaaS integrations, stealing tokens to access hundreds of customer environments
• Reports suggested up to 700 victims; attackers exfiltrated data via legitimate APIs, making activity appear normal and persistent
• Microsoft responded with Defender for Cloud Apps upgrades, adding richer telemetry, near‑real‑time detection, and stronger governance over OAuth‑connected applications

The ShinyHunters cybercrime group were so creative in breaking into corporate Salesforce environments that they forced Microsoft’s hand, making the company introduce new security upgrades just to address the attacks.

Microsoft has revealed it is focusing on improving visibility into OAuth-connected applications and strengthening governance over third-party integrations in Microsoft Defender for Cloud Apps. The changes fall into two main categories: Improved detection and investigation, and new posture and governance capabilities.

It makes sense, given that some reports claimed as many as 700 victims of the year-long campaign.

Changes and improvements

But first, a little context: In August 2025, it was reported that ShinyHunters operatives were calling their targets on the phone, claiming to be IT support, and convincing them to authorize a seemingly legitimate Salesforce Data Loader application. This app was, in fact, controlled by the attackers and requested OAuth permissions which allowed them to access Salesforce data through official APIs.

Since everything happened through legitimate authentication and API calls, the activity looked like normal user behavior.

In the following months, the campaign evolved. Instead of tricking individual employees, ShinyHunters compromised third-party SaaS providers that integrated with Salesforce, including Salesloft's Drift integration, Gainsight, and later Klue.

By stealing OAuth tokens or integration secrets from these vendors, they accessed hundreds of downstream customer Salesforce environments without interacting with each customer individually.

At one point, Google told reporters it was aware of more than 700 potentially impacted organizations.

“Microsoft consulted with Salesforce to improve granularity in telemetry for Defender for Cloud Apps with near-real-time detection, offering connected application attribution and expanded application permission insights,” the company said in a new report. “This activity was not the result of a vulnerability inherent to Salesforce. Rather, the threat actors abused trusted OAuth relationships for unauthorized access, data exfiltration, and persistence.”

In other words, Microsoft enabled greater visibility into OAuth-connected applications and their activity, allowed for better detection of suspicious API and OAuth behavior through richer telemetry and correlation, and now provides stronger governance of connected apps through permission analysis, risk scoring, and lifecycle management.

We reviewed the Bambu Lab P2S and found it "about as refined as they come" for anyone upgrading from an entry-level 3D printer.

So, I was pleased to see the Bambu Lab P2S Combo with AMS 2 Pro is $699 (was $799) at Bambu Lab — a $100 saving on a fully enclosed CoreXY printer built for multi-color, engineering-grade work. In the UK, the Bambu Lab P2S Combo also got a discount down to £619 (was £699) at Bambu Lab.

In our testing, we ran the machine constantly for weeks with very little in the way of failures or misprints, and it was clear that many of the features — the touchscreen, the AI error detection, the general polish — had trickled down from Bambu's pricier X1 Carbon and H2D machines.

Today's top Bambu Lab 3D printer deal

Scoring 4.5 stars in our review with a Highly Recommended award, we called the P2S "absolutely superb" and "about as refined as they come". It's a follow-up to the best-selling P1S, and we found the upgrades meaningful rather than just cosmetic.

The PMSM servo extruder is the standout hardware change. It delivers up to 8.5kg of maximum extrusion force — 70% more than the outgoing stepper-driven design — and samples resistance and position at 20kHz, so it can catch filament grinding or clogs in real time rather than letting a print fail silently. Combined with a high-frequency eddy current sensor for flow calibration, this printer is built to handle tricky or higher-flow materials with greater consistency than its predecessor.

The AMS 2 Pro included in this Combo bundle is worth the extra cost on its own. It's a four-spool multi-material system with active air-vent drying, sealed storage, and RFID filament sync with Bambu's own spools — useful if you print with moisture-sensitive materials like nylon or want to avoid the usual dry-box shuffle. It can double as a filament dryer even when you're not actively printing multi-color jobs.

Build quality elsewhere follows the same pattern of small, sensible refinements: hardened steel rods in place of the P1S's carbon rods, a metal build plate base, a quick-swap nozzle system that releases with a single clip, and a full enclosure with an Active Airflow system that draws in cool air for low-temp filaments like PLA, then switches to internal circulation to keep the chamber warm for higher-temp materials like ABS or ASA.

Worth considering: the P2S doesn't have an actively heated chamber, so while the enclosure holds heat reasonably well (around 50°C with a 100°C bed in our testing), it's not built for demanding high-temperature engineering filaments like PA-CF or PEEK the way Bambu's H-series machines are. Multi-color printing with the AMS still produces some filament waste from purging between color changes, and full remote-monitoring features depend on staying connected to Bambu's cloud — LAN-only mode works for printing but drops those extras.

For more top-performers, see our guide to the best 3D printers we've tested.

Also consider: More Bambu Lab 3D printer deals

• US: See all 3D printer deals at Bambu Lab
• UK: See the 3D printer sale at Bambu Lab

If you're building or upgrading a PCIe Gen 5 system, this is one of the fastest consumer drives you can buy. Right now, the 2TB WD_Black SN8100 is $400 (was $440) at Amazon, a $40 savings on a drive that tops out at nearly 15,000 MB/s.

That speed matters most for large sequential transfers — moving big game installs, video files, or disk images — where you'll actually see the difference over a Gen 4 drive.

For typical everyday use like booting Windows or opening applications, the gains over a good Gen 4 SSD are much smaller, since most everyday tasks don't come close to saturating even Gen 4 bandwidth. But if you're running video editing projects, AI applications, and games, the difference will be immediately noticeable.

Today's top WD-Black SSD deal

The SN8100 is WD's flagship consumer drive, built on SanDisk's latest PCIe Gen 5 controller and 3D NAND. Its rated 14,900MB/s read speed puts it near the top of the current consumer SSD market, and it's one of the drives we recommend as a best-overall pick for anyone with a Gen 5-capable motherboard looking to actually use that bandwidth rather than leave it on the table with an older Gen 4 drive.

This is the non-heatsink version, which is the right call if your motherboard already includes an M.2 heatsink or shroud over the Gen 5 slot (increasingly common on recent boards) — stacking a drive heatsink underneath an integrated one can trap heat rather than dissipate it.

If your board doesn't have one built in, factor in the cost of a separate M.2 heatsink, since Gen 5 drives run hot under sustained load and can thermal-throttle without adequate cooling.

One honest caveat: PCIe Gen 5 drives like this only reach their rated speeds on a motherboard and CPU platform with genuine Gen 5 M.2 support — check your board's specs before buying, since plugging this into a Gen 4 slot caps performance at Gen 4 speeds and makes the premium harder to justify.

If you have a Gen 5-ready system and want genuinely top-tier storage speed, the SN8100 is a strong buy.

For more savings, see our round-up of all the current best SSD deals.

• NSA, FBI, CISA, and 15 allied agencies warn Russia’s FSB Center 16 is exploiting weak/default credentials and old Cisco flaws to compromise critical infrastructure devices
• Advisory highlights CVE‑2018‑0171 (Smart Install DoS/RCE) and CVE‑2008‑412813 (CSRF in Cisco IOS 12.4) as examples of vulnerabilities still being abused
• TTPs overlap with Chinese groups but attribution points to Russian actors like Berserk Bear and Energetic Bear; full IoCs and mitigations were published in the joint advisory

Russian state-sponsored threat actors are continuously targeting broken and poorly configured networking devices belonging to critical infrastructure providers all around the world, a joint security advisory published by the US National Security Agency (NSA) and more than a dozen other agencies has warned.

As per the advisory, hackers working for the Russian Federal Security Service (FSB) Center 16 are constantly scanning for routers and other internet-connected devices that can be accessed with “common or default” login credentials.

Once found, these devices are instructed to copy device configuration files and later exfiltrate them via the Trivial File Transfer Protocol to servers under their control.

Berserk Bear and Salt Typhoon

In cases where default or weak credentials don’t work, the threat actors also try to exploit vulnerabilities. In the advisory, the agencies specifically mentioned two flaws in Cisco devices - CVE-2018-0171 and CVE-2008-412813. The former is an eight-year-old bug in the Smart Install feature of Cisco IOS Software and Cisco IOS XE Software that allows an unauthenticated, remote attacker to cause a denial of service (DoS) condition, or to execute arbitrary code.

The latter is an even older (18 years old) set of multiple cross-site request forgery (CSRF) vulnerabilities in the HTTP Administration component in Cisco IOS 12.4 on the 871 Integrated Services Router that allows remote attackers to execute arbitrary commands.

Even though many of these tactics, techniques, and procedures (TTP) overlap with Chinese hackers Salt Typhoon, the agencies suggested they are primarily focusing on Russian hackers known as Berserk Bear, Energetic Bear, Crouching Yeti, Dragonfly, Ghost Blizzard, or Static Tundra.

The joint advisory is co-authored by the NSA, FBI, and CISA, as well as 15 other agencies from Australia, the United Kingdom, Canada, New Zealand, Estonia, Finland, France, and Italy.

Artificial intelligence has quickly become the boardroom's favorite solution. From forecasting and reporting to scenario planning and budgeting, finance leaders are under growing pressure to demonstrate how AI can improve efficiency and drive better decisions.

But against the rush to adopt AI, many organizations are overlooking a fundamental truth: that AI is only as effective as the systems, processes and data that support it.

This is particularly true in finance, where many teams continue to rely on fragmented technology stacks, disconnected data sources and spreadsheet-heavy workflows. While AI promises to automate analysis and surface deeper insights, it cannot compensate for weak foundations. In fact, it often does the opposite, exposing issues that previously remained hidden beneath layers of manual work.

The reality is that many finance functions are less prepared for AI than they realize.

The spreadsheet problem AI cannot solve:

Spreadsheets remain deeply embedded within enterprise finance. They are familiar, flexible and accessible. However, they were never designed to serve as the backbone of modern financial planning and analysis for large enterprises.

In many organizations, critical forecasting models, budgeting processes and reporting workflows are still maintained across countless spreadsheets, often with limited governance and varying levels of accuracy. Data is copied between systems, formulas evolve over time, and key assumptions can become difficult to trace.

Introducing AI into this environment does not eliminate these challenges. It amplifies them.

If an AI model is drawing insights from inconsistent data sources or outdated spreadsheets, it will simply generate the wrong answer faster. Automated recommendations may appear sophisticated, but their reliability is ultimately determined by the quality and integrity of the underlying information.

This is why the familiar principle of ‘garbage in, garbage out’ remains so relevant to finance teams today.

Why finance teams may be overestimating their AI readiness:

Many organizations assess AI readiness by evaluating tools. They ask whether they have access to the latest models, whether employees are using generative AI and AI agents, or whether automation opportunities exist within their workflows.

Far fewer assess the quality of the infrastructure feeding those systems.