The Register

BORK!BORK!BORK! Europe's airports can be a minefield for the unwary traveler, especially when that traveler inserts a passport into a waiting kiosk and finds Windows staring back. An eagle-eyed Register reader (they said we could use their name, but we will keep them anonymous for fear of them being added to some sort of list) encountered the Windows desktop where it probably didn't belong this month, while traveling through that little-known French provincial airport: Paris-Charles de Gaulle. Our reader was making their way through the airport after returning from holiday. They told us, "I was about to go through a PARAFE gate (the French equivalent of the British ePassport gates) when suddenly, right after I inserted my passport, the system crashed and rebooted." Where there had been an instruction to insert a passport and hope you looked enough like your picture that a Gendarme didn't show up with a rubber glove and an invitation to the special room, there was now a Windows desktop (Windows 11, by our reckoning) and a closed gate. Eventually, somebody launched the appropriate application (not in person, our reader told us), and the queue could shuffle forward again. Paris-Charles de Gaulle is a special airport. While our reader had the Windows experience in Terminal 3, this writer missed an April flight to the UK from Terminal 2 after another computer whoopsie at passport control caused an impressive delay. At the time, an airport spokesperson said, "We can confirm these days, the police encountered IT issues affecting their passport verification system, which may have caused significant delays in passenger processing." A great comfort, as a certain airline with orange livery shut up shop for the night. After all, it wasn't their systems that had succumbed to a wave from Le Bork fairy's wand. We've noted a bit of a run on airport borks of late. Sometimes signage shows things it shouldn't, or boards indicate that cost-cutting has perhaps gone too far. The curse of bork spreading to border control, however, is something new. But it remains reassuring to remember that, even as security is tightened and regulations made more onerous, Windows can still be guaranteed to pop up in the most unexpected places. ®

The new "bullet-proof" KDE Software Initiative brings long-term fixes to Kubuntu – and other distros can take advantage of them too. Announced late last week, the effort aims to provide "at least" three years of backported fixes for KDE Plasma 6.6 and the associated KDE Frameworks and KDE Gear apps. It is aimed at Kubuntu 26.04, the latest LTS release of Ubuntu's official KDE Plasma flavor, but other distros are free to pick up the maintained components. This effectively makes Plasma 6.6, which debuted six months ago, a new LTS version of the desktop. As we reported back in May, Kubuntu 26.04 came with Plasma 6.6 – specifically, version 6.6.4. By July, that had been superseded by Plasma 6.6.6, which, as we said at the time, was intended to be the final release in the Plasma 6.6 cycle. As KDE developer Nate Graham explains in a blog post titled What a real LTS looks like: Kubuntu 26.04, the driving force behind the new LTS effort is that an organization called Kubuntu Focus stepped up to pay for it. It's sponsoring Graham's company TechPaladin – which we mentioned last year – to do the maintenance work. The arrangement means that fixes will cover not only the desktop itself, but also its underlying code libraries and accompanying suite of apps. Specifically, that means version 6.24 of the KDE Frameworks, plus version 25.12 of the KDE Gear application suite. The primary beneficiary is Kubuntu, specifically version 26.04, but other operating systems and Linux distros are free to incorporate the updates. The three-year support lifespan is significant here, and we feel that we need to spell out why. Although there are ten official Ubuntu flavors, an important detail that's often overlooked is that only the official GNOME version of Ubuntu gets the full five-year support lifespan. The other flavors do not share that long support lifespan. Beneath their shiny graphical shells, the text-mode stuff – the part most GUI users never see – is the same Ubuntu OS. That gets updates for the full LTS lifecycle. The graphical desktop on top doesn't. Kubuntu is the oldest of the remixes, but even so, it only gets three years of support, not five. It's right there in the release notes: "Kubuntu 26.04 LTS will receive security updates and critical fixes through April 2029." Compare and contrast with the GNOME edition: as its release notes say, "Ubuntu 26.04 LTS will be supported until April 2031." Once the desktop version bundled with an Ubuntu LTS is superseded, users who want updates to the GUI components must upgrade them separately. As an example, back in late 2022, we covered the release of new versions of KDE and LXQt for Ubuntu 22.04. By fortunate coincidence, Kubuntu 24.04 shipped with Plasma 5.27, the final release series of Plasma 5. (Specifically, it had 5.27.11, although a final version 5.27.12 did appear the following January.) Now that a backer is funding the security effort, the KDE team can maintain the entire desktop stack more comprehensively. So long as other distributors pick this up, this should be good news for anyone using KDE on a long-term distro or OS. It's also good for Kubuntu, which faces growing competition. Although it's far from finished yet, there's now an official KDE Linux distro. We described last year how KDE Linux is based on Arch and has robustness and fault-tolerance features similar to those in Valve's Steam OS 3. The new OS complements the older KDE Neon distro, and may in time replace it. KDE Neon is a far more conventional distro: like Kubuntu, it's based on Ubuntu LTS releases, but it's constantly changing as new builds of KDE appear. It's not ideal for long-term deployment. Another contender is Tuxedo Computer's Tuxedo OS. We looked at Tuxedo OS 1 back in 2022, but recently the company has announced major changes. Early last month, the company announced that it would switch from an Ubuntu base to Debian – but, unusually, use the rolling Testing branch rather than Debian Stable. The company cited difficulties with keeping Snap packaging out of its remix, as well as Canonical's new AI focus. As of this weekend, a beta release of the new version is available. ®

Britain is putting Google AI in the flight-planning loop to see whether airliners can dodge the patches of sky where their vapor trails are most likely to stick around and warm the planet. The UK government is backing a trial to predict where persistent contrails will form over the North Atlantic and help participating aircraft avoid those areas by adjusting their routes or altitude. The Department for Transport is putting £2.6 million into the 30-month project, which will combine Met Office weather forecasts with Google's artificial intelligence to suggest alternative flight paths. Flight tests are scheduled for the winters of 2026-27 and 2027-28. Any route or altitude changes will go through normal air traffic control procedures and require the usual safety checks. Those white streaks behind aircraft aren't just an opportunity for conspiracy theorists to get excited. Contrails form when water vapor from jet exhaust freezes into ice crystals in cold, humid air. Many disappear quickly, but under the right atmospheric conditions they can hang around, spread out and develop into clouds that trap heat. Persistent contrails are therefore one of aviation's non-CO2 contributions to climate change – and potentially one that can be tackled without waiting for airlines to replace their fleets or invent a jet engine that runs on good intentions. The trial rests on the fact that contrail-friendly conditions occur only in particular regions and layers of the atmosphere. If forecasts can identify them accurately enough, aircraft could fly above, below, or around them. "We're partnering with Google to back British experts and innovators to find practical ways to make flying cleaner," said aviation minister Keir Mather. "By testing small tweaks to flight paths over the Atlantic, we can cut the vapour trails left behind by planes." There is, naturally, a trade-off. Changing an aircraft's route or altitude can burn additional fuel, potentially increasing its CO2 emissions even as it avoids a contrail. The trial will therefore need to establish whether the climate benefit of dodging persistent trails outweighs any extra fuel burned getting out of their way. The project joins a much larger pile of taxpayer cash being pointed at greener aviation. The government announced £219 million for sustainable aviation fuel production in June and is spending another £43 million on technologies including electric and hydrogen-powered aircraft. This experiment, though, does not require anyone to reinvent the airplane. The government and Google are betting that sometimes a greener flight path is only a few thousand feet away. ®

The process of training an AI model becomes a paradox at scale – the more it remembers, the less it remembers about the source of its memories. MIT computer scientists went looking for a way to attribute AI model output to specific training data, in the hope that understanding could inform AI regulation. What they found, described in a paper titled, "Outputs of Generative Diffusion Models are Often Unattributable," looks like it will actually make regulation more difficult. Scientific journal Nature Communications will publish the paper on Tuesday. The authors, Zheng Dai and David K Gifford, affiliated with MIT's Computer Science & Artificial Intelligence Laboratory (CSAIL), note that diffusion models like Midjourney and Stable Diffusion have become widely used tools for generating artifacts including images, videos, and audio. Diffusion models have also attracted lawsuits from artists who argue that copies of their work included in training data have enabled AI models to reproduce their output and artistic style. In one such ongoing copyright case from 2023, Andersen et al. v. Stability AI Ltd, the plaintiffs have been trying to convince the court to make defendant Midjourney provide the datasets used to train its models. The plaintiffs allege that Midjourney made its training datasets "by scraping images associated with specific artists’ names for the express purpose of enabling its model to mimic those artists’ expressive content." Being able to attribute model output to the content they ingested during training would help people understand how models function and would have various applications "including machine unlearning, data poisoning, model interpretability, fairness, and privacy," Dai and Gifford wrote in their paper. "Furthermore, given the contemporary adoption of these models for creative and commercial purposes, attributability also carries ethical, policy, financial, and legal implications." But as it turns out, attributing model output to a specific input becomes more difficult as models get larger. "Here we show that attribution, characterized as the task of locating a part of the training data that can be held responsible for a generated sample, can become impossible if a model is trained on a sufficiently large corpus of data," the authors state. "We find that the more data a model is trained on, the less attributable its generated samples become, a phenomenon we henceforth refer to as attribution decay." The authors tested this by removing specific training data through a process called ablation. And the result of this testing showed that for very large models, you could take away, for example, the image of the Mona Lisa or all of Leonardo Da Vinci's work — yet the model could still reproduce that image or style. In MIT's press release, Gifford argues that the findings suggest models are creative in the sense that they're not just copying their training data. "If those outputs have nothing to do with any individual piece of training data, that raises questions about fair use, about whether the outputs are themselves copyrightable as novel works, and about how authors get compensated when what comes out of a model isn't attributable to anything on the internet," he said. He also argues that having the ability to test whether a given output is derivative creates an obligation for companies to show their work cannot be attributed to a particular source. At the same time, the research suggests a liability avoidance strategy – make your model large enough that no output can be attributed to any one specific input. James Grimmelmann, a law professor at Cornell Law School and Cornell Tech, said in a statement, “If attribution worked, it would reliably tell us whether similarities between a model's output and a copyright-protected work are due to copying or coincidence. But this paper provides reason to think that attribution will fail for interesting models. Instead, technologists and courts will need to resort to other methods for assessing copying." In an email to The Register, Grimmelmann said current copyright claims against AI companies haven't focused specifically on the extent to which images similar to artist’s real-world work can be elicited from a model. "German cases have dealt with apparently memorized outputs from music models, and US cases have dealt with whether training is a fair use, but output similarity for image models remains mostly untested in court," he said. ®

Fewer skilled overseas workers are applying for UK tech visas, with numbers down for the third year running even as ministers talk up Britain's chances in the global technology race. The tally fell 7 percent, from 37,376 applications in 2024 to 34,936 last year, according to UK Visas and Immigration figures obtained by accounting company RSM UK and shared with The Register. That's a long way from the 53,729 applications recorded in 2022, when the numbers peaked. RSM reckons the decline risks making an already difficult hunt for tech talent even harder. The underlying data also shows which occupations accounted for most of the overseas tech recruitment. Programmers and software developers were by far the largest group in 2025, with 15,689 certificates, followed by IT business analysts, architects, and systems designers with 11,821. IT quality and testing professionals accounted for another 3,091, cybersecurity professionals for 858, and IT network professionals for 310. Comparisons between individual occupations are complicated by a government classification change in April 2024, which split that year's figures between old and new codes. The overall total was unaffected. RSM nevertheless argues that the trend should concern an industry already struggling to recruit people with specialist skills. "Talent shortage remains one of the biggest challenges in the tech industry, and risks holding back growth, both short-term and in the long run, without some form of intervention," said Ben Bilsland, partner and head of technology industry at RSM UK. Bilsland said higher employment costs had made businesses wary of hiring and warned that AI wasn't going to magic the problem away, since companies still need skilled staff to deploy the technology "safely and successfully." RSM is calling for visa reform, pointing to comments from UK AI minister Kanishka Narayan about attracting more top engineers to Britain and making it easier for technology companies to scale at home. Bilsland said any reforms would need to be fast-tracked, adding that the application process should be accessible to tech companies of all sizes. Britain isn't alone in seeing fewer overseas workers knocking at the door. Across the Atlantic, H-1B registrations for the US government's 2026 fiscal year fell about 25 percent year-on-year, as The Register reported last year. US tech companies rely heavily on the visa to recruit foreign workers for specialist roles. There was more going on in the US, however. Immigration officials had introduced measures aimed at curbing duplicate registrations and fraud, including a one-entry-per-beneficiary system, while the registration fee jumped from $10 to $215. Still, both sets of numbers land as governments compete to attract the engineers, developers, and other specialists expected to build the next generation of AI and other emerging technologies. ®

Google has acquired the data of failed US airline Spirit. Spirit hit financial turbulence when COVID-19 blew in during 2020 and started making losses. Its balance sheet never climbed back to a safe altitude and in May 2026 the airline grounded itself permanently. The low-cost carrier entered liquidation to wind itself up and is now auctioning assets to raise cash and settle at least some of its debts. A court document [PDF] filed last week reveals that one of the assets up for sale is a huge trove of deidentified data, which Google bid for and won for just $10 million. For that sum, Google bought itself 100 million emails and 500 million items from Microsoft Teams, 17 million OneDrive files and 20.5 million items from SharePoint. The search giant also now owns over 30 million recorded customer service calls, and more than 15 million customer service chat records. 600,000 ServiceNow tickets are another element of the collection, along with 13.7 million active emails addresses from Oracle’s Responsys marketing application, and details of 11 million sales of in-flight Wi-Fi services. There’s also operational data in the trove, describing over 763,000 flights, five million crew pairings, more than 1.2 million fuel slips, and records describing purchases of 787,452 parts. Google has reportedly said it bought the data to improve its AI services. The underbidder was Mercor, a company that provides data to train AI models. So clearly Spirit’s data is of value to AI companies. Indeed, The Register recently reported how AI experts increasingly believe large language models are blunt instruments, and that smaller models trained on specific fields of knowledge are more useful in some applications. Google might have itself the basis for an aviation ops model, or with all sorts of quotidian financial records that could be useful for another AI. If you’ve flown Spirit and worry that Google will soon know about a testy conversation you had with the airline’s call center, you’re being told not to worry. The court filing says the data was deidentified before being put on sale and Google has promised to scrub any PII it finds in the trove. The Register awaits evidence of the inevitable errors that mean some personal info appears as the result of a future prompt, or search. Fasten your seat belts! ®

If you’re struggling to stay up to date with Redmond’s regular product re-branding exercises, here’s the site for you: The Microsoft Rebrand Registry. Microsoft Most Valuable Professional (MVP) Loryan Strant created the site because he thinks it’s a valuable resource, and also in the hope it makes visitors “chuckle.” Readers may remember that Strant has also created the site Let Me Correct That For You, which lists the exact names of Microsoft products – an effort he told The Register he thinks is useful because Microsoft in its wisdom uses Camel Case for names like PowerPoint but went with conventional capitalization for Copilot. Another of his sites immortalizes Microsoft cloud product logos. He’s also created HumbledandHonored.com, a site that generates social media posts MVPs can use to announce they have earned or retained Microsoft’s awards. Strant told The Register that the Rebrand Registry came about after some banter between himself and other MVPs, during which the topic of Microsoft’s many product name changes came up. He decided to do something about it. The site lists 72 Microsoft products and 158 names they’ve had over the years. It also includes some analysis of Microsoft’s branding, which sees a product’s name survive for an average of two years and eleven months. The site lists eight products that have gone through three name changes: · Azure AI Search · Azure App Service · Azure DevOps · Azure SQL Database · Foundry Tools · Microsoft 365 Copilot app · Microsoft Configuration Manager · Microsoft Defender for Endpoint Strant has even had a go at predicting which Microsoft products are likely to get a new name soon, by considering the amount of time the current name has applied, prior names, and the frequency with which Microsoft changes names of products in the same family. That methodology led him to suggest an “elevated” likelihood of name changes for the Azure App Service, Azure SQL Database, Azure DevOps, and Microsoft Dynamics 365 Field Service. The site considers surviving products only, and Strant admits his analysis is therefore biased towards product Microsoft still publishes. He also pointed out that he’s not decided how to handle successor products, for example when Microsoft discontinued Skype for Business and put similar functions into Teams. Strant said he created Let Me Correct That For You using WordPress, while his Microsoft logo library grew out of a OneDrive folder and relies on a GitHub repo. For the Rebrand Registry, he used vibe coding tools. “I have my own harness that lets me custom-build things,” he explained, adding that he rigorously validates data rather than relying on AI tools to get facts right. Strant said most of the feedback he gets about his sites is positive, either on LinkedIn or during face-to-face meetings at events. He’s heard that Microsoft people appreciate his efforts because Redmond hasn’t preserved the same data his sites record. “I hope that people find them useful, and get a chuckle,” he said. ®

The Xen Project, home of the open-source hypervisor, has decided it’s time to get serious about formal safety. That’s not an admission of security woes, but a push to ensure the hypervisor complies with formal safety standards like IEC-61508, which defines the processes and practices used in software that runs in devices and settings that could conceivably harm humans. Xen has in recent years emphasized work to make its code more suitable to run in embedded workloads and devices like cars. The project points to the fact that modern vehicles rely on software for tasks as diverse as emergency braking and infotainment systems, and manufacturers’ preference to offer all such services with as few physical computers as possible. Virtualizing workloads in a car means a glitchy navigation app is isolated and can’t impact performance of a critical application like collision avoidance. The Xen Project now wants to make its software an even better option in such scenarios by complying with formal safety standards. The Register understands that project members believe many more devices, such as robots, will also need workload isolation, so they want to make sure that Xen becomes hypervisor-of-choice in more scenarios. The group has therefore created a Safety Committee. Three companies are aboard as founding contributors – AMD, EPAM, and Renesas. AMD and Renesas are chipmakers who are always on the lookout for new markets. Working with Xen and demonstrating expertise in a safety-certified hypervisor won’t hurt them one bit as they pitch their silicon to manufacturers. The Xen Project has also created a new membership tier called Premier Plus for organizations pursuing functional safety certifications. “Safety engineering has traditionally required every organization to recreate much of the same foundational work. We believe open source can change that,” said Cody Zuschlag, community manager at Xen Project. “By maintaining shared engineering evidence alongside the software itself, the Xen community can reduce duplicated effort, accelerate certification programs, and make functional safety more practical for industries building the next generation of software-defined systems.” Boeing recently joined the project, and The Reg understands one reason the aerospace giant signed up was its interest in Xen's safety work. It’s been a busy couple of weeks for the Xen Project, which in early August delivered an update to its flagship hypervisor. It wasn’t a major release but did add support for AMD Zen 5 Bus Lock Threshold, which makes it easier to ensure one guest VM doesn’t make it harder for other workloads to access CPU resources. The release also improved Arm support, and advanced work towards a fully functional release for RISC-V processors. ®

Are web coders losing their respect for standards? A recent review of the world’s largest sites suggests so. ValidateHTML, a project by independent French developer Théo Ducreux, examined the contents of the 5,000 most widely used web domains, as estimated by the Tranco research project. They include the usual suspects (Google, YouTube, Akamai) plus those without public-facing components (Google’s “gstatic,” EZVIZ’s “Ezviz7”). Ducreux found more than half – 2,656 sites – served a home page readable by humans. Many of the others, such as Google and EZVIZ’s, were just traffic redirection sites, a beacon for a company’s products to call home. Headless or not, nearly 90 percent of these sites were not serving proper HTML, as specified by the World Wide Web Consortium (W3C) and the WHATWG (Web Hypertext Application Technology Working Group). These standards are important so that different browsers and reading tools can render the requested pages in more or less the same way. Ducreux tallied a total of 100,305 HTML violations (including the most recent HTML5 specs) across all these sites, plus an additional 18,863 CSS errors. In other words, 87.2 percent of sites violate web specs in at least one place. Only 12.8 percent have fully valid HTML. And only 2.6 percent of all sites are completely clean, returning zero errors and zero best-practice warnings. More alarming, over a third of sites failed accessibility checks, which to their owners is more of a legal liability (and hence a compliance issue) than an embarrassment, given the potential reach of the 2025 European Accessibility Act. (Here in the U.S., woe be the accessibility-inconsiderate e-commerce site that raises the ire of the National Federation of the Blind.) Accessibility-wise, 20.4 percent of sites are missing alt text on images, which often contain vital information. Worse, 41.6 percent of pages are missing ARIA labels that identify page regions, potentially leaving screen readers struggling to convey the page's structure. Here’s an iframe, figure it out. “Screen readers don't have anywhere near the error tolerance Chrome has, so markup that looks totally fine to you can be broken for someone using assistive tech,” Ducreux told The Register. For the project, he used a home-built web crawler and a set of open-source parsers (HTML-validate, CSS Validator, Lightning CSS, fast-xml-parser). “The site, the scoring, and the crawl logic are my own code,” Ducreux explained by email. Ducreux is the principal maintainer for the project and the site. "No team, no funding, no company." An electronic Tower of Babel Does lack of adherence to standards really matter? It's an open question. “The web is more a social creation than a technical one,” web creator Tim Berners-Lee wrote back in 1999. If we all go off and make our own standards, then it's harder to connect and communicate. The currently fashionable Dead Internet Theory – in which AI machines and not we meat sticks now make up most internet traffic – won’t get developers off the hook. As Shopify engineers have found, even superintelligence prefers shipshape markup. Browsers deserve part of the blame. Unless they have a fetish for the short-lived but brutally unforgiving XHTML, browsers today will just ignore any code they can’t grok. “Browsers are so good at error recovery that nothing forces anyone to fix” their code, Ducreux wrote on the site, “which is exactly why it accumulates.” Indulgence encourages slothfulness. Front-end frameworks create their own mischief. The single most common failure identified (in more than 59 percent of all the sites) is the misplaced element, aka incorrect tag nesting. Did you know you're not allowed to put a style element within a div box? Many frameworks do, evidently. “It is a nesting problem created at build time, not in an editor,” Ducreux wrote on the site. “Nobody writes that by hand.” Many folks believe sloppy HTML is no big deal, including some in the peanut gallery at Hacker News. One contributor noted that “HTML5 defines a method for turning more-or-less any sequence of bytes into the same DOM tree,” DOM being the Document Object Model the browser uses to understand a web page. Ducreux himself doesn’t see bad HTML as an existential threat to the web, outside of hindering accessibility, which can leave someone guessing at online content they may really need to understand. “Honestly, most invalid HTML doesn't break anything visible, browsers are built to guess what you meant and just patch it silently,” he wrote by e-mail. But at the same time, the web isn’t just about browsers. “AI agents reading pages, voice assistants, translation tools, screen readers – none of them have Chrome's 20 years of guessing your intent,” he wrote. “Writing to spec is what makes a page mean the same thing to all of them, not just to whoever's rendering it in Chrome that day.” ®

Payments biz Stripe has reportedly finalized an agreement to acquire OpenRouter for at least $7 billion, signaling a shift in focus for the AI industry. OpenRouter provides a way for customers to integrate multiple AI models into their workflow and infrastructure and is, by Ramp's measure, the most popular of the gateway companies that have sprung up to simplify model integration. By acquiring the AI gateway service, Stripe would become a mediator of AI token sales, fitting its core business while gaining access to valuable data about AI model usage. Stripe declined to comment, citing a policy of not addressing rumors or speculation. Prior to Bloomberg's report about the agreement, the Wall Street Journal said acquisition talks were underway. Akhil Verghese, founder and CEO of Krazimo, a developer of AI software for businesses, told The Register in an email that he found the acquisition fascinating because OpenRouter is to AI models what Stripe has been to financial infrastructure. "A key factor for me is the value," he said. "I won’t pretend to be smart enough to understand how these valuations work, but $7 billion for OpenRouter surprised me. They raised at a $1.3B post money valuation in May, and even if revenue has grown significantly since then and acquisitions often have an additional multiple, 5x in 3 months is nuts. "There are ways this acquisition makes sense, though. [Stripe CEO] Patrick Collison recently said that metered pricing is the native business model of the AI era. In purchasing OpenRouter, Stripe, which already controls where the money is going, now sees where the tokens are going." And there are a lot of tokens going around. An economics paper published last month, AI Premium, estimates total global LLM token consumption to be about 5 to 7 quadrillion per month, with about 2 percent of that being handled by OpenRouter. While marquee model makers Anthropic and OpenAI have focused on building their own brands and steering customers toward their tooling, the reality is that business customers prefer options that avoid lock-in and maintain negotiating power. That expectation and the potential fees for gatekeeping have led to the proliferation of AI gateway firms and to $113 million in funding for OpenRouter back in May. Verghese said rival AI gateways may find it more difficult to compete now that OpenRouter has access to so much funding. Anthropic and OpenAI account for a substantial amount of the tokens processed by virtue of their subsidized flat-rate subscription plans – generally served directly rather than through a gateway or third-party provider. But as they push customers toward API pricing in preparation for going public, they risk driving business toward providers of more affordable, open weight models. As of December 2025, OpenRouter reported serving more than 5 million developers to route traffic to more than 300 models from more than 70 providers. At the time, open weight models accounted for about 30 percent of tokens served. Today, the biz boasts more than 10 million developers and 80 providers. OpenAI in October 2025 was serving about 8.6 trillion tokens per day, according to Andreessen Horowitz, and OpenRouter was serving more than 1 trillion tokens per day. "The competitive frontier is no longer only about accuracy or benchmarks," said Malika Aubakirova and Anjney Midha from Andreessen Horowitz at the time. "It is about orchestration, control, and a model’s ability to operate as a reliable agent." It's about AI gateways and adjacent services that sit above models in the technical stack. According to Vercel, which offers its own AI Gateway, open weight models have seen their share of gateway token spend grow from 11 percent in April to 36 percent in July. Meanwhile, the token spend collected by the four largest frontier labs, which had not dipped below 93 percent in seven months, fell to 89 percent last month. Much of that came at Google's expense, with Chinese AI labs Z.ai and Moonshot capturing the growth in open weight spending. "GLM 5.2 and Kimi K3, both released in the last two months, are the first open-weight models running a meaningful share of the workloads historically owned by closed-weight labs," Vercel said. Even so, the top four frontier labs in the US continue to capture the lion's share of spending. Vercel said those companies took 95 percent of spending through its AI Gateway in June. "In July, Anthropic collected 65 percent of gateway spending on 30 percent of token volume, at 4.4 times the average price of every other lab's tokens," Vercel said. Open weight models, which have occupied the low end of the market (cheap tokens), have seen their fortunes shift, Vercel said. After their share of gateway token volume almost tripled between April and June, from 11 percent to 29 percent, capturing only four cents for every gateway dollar, they saw spending more than double to nearly nine cents for every dollar. Open weight models, in other words, are becoming better and are bringing in more revenue even as the average token price declined 13.6 percent in July. If the trend continues and AI usage drifts further toward open weight models and away from proprietary frontier models, the AI gateway business should prosper. Verghese expressed concern about whether OpenRouter will be able to maintain its neutrality under Stripe's roof. The workload router isn't supposed to play favorites in terms of where it directs tokens based on requirements, but should it start doing so, he said, the resulting loss of trust could be harmful. ®

Siemens and fellow German engineering biz Reinhausen are developing a solid-state transformer (SST) to deliver 800-volt DC power directly to increasingly dense AI datacenter racks. As AI hardware grows hotter and hungrier, server farm operators are looking to overhaul how power moves through their data halls. One approach, favored by Nvidia and other industry heavyweights, is distribution at 800 VDC. The two electrical enterprises are therefore developing a modular SST capable of converting medium-voltage AC at up to 36 kV into a stable 800 VDC output for downstream distribution. Siemens says that the SST will provide a compact, galvanically isolated alternative to conventional multistage AC-to-DC power conversion architectures, reducing intermediate conversion stages and enabling more efficient delivery to high-density AI infrastructure. According to Reinhausen, datacenters supporting AI workloads are among the industry's most demanding power consumers. Sudden load spikes can strain servers, storage, and cooling systems, reducing performance and potentially causing outages. "Our technology enables AI datacenters to operate with the efficiency and reliability they require," said Reinhausen managing director Wilfried Breuer. "Solid-state transformers are critical for these large and power intensive structures. They enable higher efficiency, direct grid-to-rack DC conversion, and a reduced footprint. This collaboration will further strengthen our leading position in the electrification of datacenters," commented Siemens CEO of Electrification and Automation, Stephan May. As The Register reported late last year, Nvidia sees 800 VDC power distribution as the way forward for AI infrastructure, and says the shift will be necessary to support IT racks drawing 1 MW or more. That may sound far-fetched when average rack density has only just passed 11 kW, according to Uptime Institute, but hyperscalers such as Google are already planning for megawatt-scale racks. Nvidia has also developed its own MGX-compatible 800 VDC power rack, which it says will arrive "in the second half of 2026." Siemens and Reinhausen said that they intend to industrialize their datacenter SST, but did not say when it would become commercially available. We asked for clarification. Another German electrical firm, Infineon, is also backing 800 VDC and positioning its silicon carbide (SiC) power semiconductors for use in the infrastructure. ®

An AI broke Snowflake’s code; then another AI, an attack agent, autonomously found the bug, exploited it, and extracted credentials without human intervention. Luckily, this wasn’t yet another case of rogue AI agents doing evil things. It was a sanctioned bug hunt, conducted through Snowflake’s HackerOne vulnerability disclosure program, and Snowflake fixed the flaw the same day Wiz reported it and rotated the affected credentials the following day. Wiz’s red agent, an AI-powered autonomous attacker designed for offensive security, found the GitHub Actions workflow flaw during a routine scan of public repositories on June 23. The script injection vulnerability existed in snowflakedb/snowflake-connector-net, and it allowed an unauthenticated user to execute arbitrary commands within a GitHub Actions runner by opening a GitHub issue with a specially crafted title. And it turned out an AI had inadvertently injected the bug into the code five days earlier. GitHub Copilot Autofix, an AI coding assistant, co-authored the commit on June 18, and it introduced a script injection bug in run: blocks by removing the repository’s existing sanitized input pattern and replacing it with direct string expansion in a shell script. “We crafted an issue title that, after template expansion, breaks out of the echo string and exfiltrates the Jira credentials via an out-of-band callback,” Wiz’s head of threat exposure Gal Nagli said in a Monday blog. These credentials gave Wiz read access to Snowflake’s engineering, security compliance, and bug bounty tracking projects. Wiz reported the workflow vulnerability to the cloud data platform on June 23, and Snowflake patched it the same day. It also revoked and rotated the Jira token, and confirmed, via audit logs, that Wiz was the only third-party to access the endpoint during the five-day exposure window. The disclosure “was immediately investigated and remediated, and our investigation found no evidence of unauthorized access,” a Snowflake spokesperson told The Register. “We are working together with Wiz to share these learnings with the broader industry to encourage widespread adoption of these security best practices.” Wiz, for its part, deleted all of the data it accessed during the vulnerability research and proof-of-concept exploit testing, and told us that this incident proves human code review isn’t sufficient to quickly detect vulnerabilities - especially as developers increasingly use AI. “This incident highlights a rapidly emerging reality in software development: how AI coding assistants can inadvertently introduce workflow injection vulnerabilities, and how automated AI agents can rapidly surface them in the wild,” Nagli wrote. Of course, the Google-owned biz has a vested interest in saying this. But this doesn’t make it not true.®

Firefox for iOS now includes an optional ad blocker, or it will soon. According to Mozilla, the experimental feature "is being introduced to the Firefox user base through a progressive rollout." As such, it might not have turned up for every user just yet. Off by default, the block uses a filter list based on EasyList to block ads before they load. According to the organization, this means that it covers third-party advertising networks, ad-related trackers, ads served up by websites, as well as pop-ups and overlays. It does not block ads that appear in search results (including Google and Bing), nor does it block "sponsored content" on Firefox Home or the New Tab Page. In addition, the company stated, "Some websites serve ads in ways that are not covered by the current filter list." Users might therefore find it less effective than alternatives such as Wipr, Ghostery, or other solutions. However, it is undeniably convenient. Hit a toggle in the app to enable or disable ad blocking, and that's it. Firefox is hardly the first to the party when it comes to browser-based ad blocking on iOS. Brave, for example, boasts all manner of privacy services and ad-blocking features. The omissions from Firefox's block list will also worry users. Far be it from us to suggest that the substantial chunk of the company's revenue coming from Google might affect what gets blocked, but the decision not to block ads from search engine results might raise an eyebrow. The company states in its FAQs, "Firefox blocks many third-party ads and advertising infrastructure, but not every advertisement can be blocked. Some websites serve ads in ways that are not covered by the current filter list." Ad blocking remains a contentious issue. The attempts to lock down Chrome's browser extension platform could be regarded as only a skirmish in an ongoing war, particularly as another more insidious front has opened up thanks to AI. A report published today by Adblock Plus warned that advertising in conversational AI is likely on the way. "Such ads," the company said, "change how ad delivery works, can be generated on the fly, and are far harder for users to recognize." And will likely make Mozilla's foray into iOS ad blocking seem like something from a gentler time. ®

Microsoft's GitHub source shack is reporting a 50 percent error rate on repository content downloads. GitHub Copilot is also "experiencing degraded availability." The problems began at 1340 UTC, when the company's status page reported "impacted performance for some GitHub services." This rapidly escalated to degraded performance across numerous services, including Issues – which was, appropriately enough, experiencing an issue of its own. GitHub is investigating and has promised an update, but today's incident is only the latest in a string of disruptions that have left developer nerves frayed. Earlier in August, GitHub's Actions automation platform and Pages hosting service experienced problems. In its May 2026 availability report, the company acknowledged that AI-assisted coding and agentic workflows were adding to the strain. GitHub parent Microsoft boasted last year that AI was writing as much as 30 percent of the code in some of its repositories, subject to human review, according to CEO Satya Nadella. Addressing GitHub's repeated outages in June, software engineering SVP Jakub Oleksy said structural changes were afoot to "permanently remove failure modes." "We acknowledge that we have work to do, but we're committed to getting it done and making GitHub reliable when and where you need it." Brave words, but today's disruption demonstrates that there is still plenty of work to do – and each recurrence further tests developer confidence in the service. Unsurprisingly, developers took to social media to vent their frustration as Monday Mondayed even harder thanks to the issues. One posted, "Github down again, abou[t] as surprising as the sun rising." Another was more plaintive: "Github is really falling apart. Can we stop being down please... We got work to do." The cause remains unknown, but GitHub's disruptions are stacking up, leaving developers to count the idle hours and consider alternatives. "Maybe," pondered one developer, "the AI code submissions are too massive for them servers." Maybe they are. ®

Do you love lasers and hate mosquitoes? For $1,000, you can install your very own bug-zapping surface-to-air missile rig at home – no military-grade budget required, provided the startup behind it delivers on its promises. The Chinese company behind Photon Matrix has opened pre-orders for its consumer-grade mosquito zappers after videos of early prototypes began circulating in July 2025. The devices resemble home theater projectors in both size and appearance, and can be deployed wherever their owner has an acute mosquito problem. They use LiDAR to locate the pests, a galvo system to track their movement, and a laser to zap the bug in mid-air, according to the vendor. Photon Matrix Lab, based in Changzhou, China, markets the device as a chemical-free alternative to mosquito repellents and insecticides. The company said the device is intended for use inside the home, including bedrooms and nurseries. Official images show it sitting on a nightstand or mounted on the wall above a bed. According to the marketing materials, mosquito haters can choose between a visible blue laser and an invisible infrared beam to take out the bloodsuckers, and the devices have an effective range of six meters. In addition to mosquitoes, the device is supposedly capable of taking out other winged irritants measuring between 2 mm and 20 mm, provided they fly no faster than 1 meter per second. You can watch it in action below. Photon Matrix can now be pre-ordered in the US, UK excluding Northern Ireland, EU, Australia, and China, with the website promising shipment within 120 days. It has taken the company more than a year to get to this stage. It first documented a working prototype in June 2025 and, after a lengthy period of R&D and more than $2.8 million raised via Indiegogo, the devices have entered mass production, according to the vendor. However, this lengthy period of development has led to unease among backers, as has the lack of assurances that the device can meet safety standards in the markets the company intends to enter. Photon Matrix Lab's most recent update on laser safety compliance (and its third most recent update overall) came in June 2026, when it said certification applications remained in progress. "All mandatory safety and quality certifications for the laser mosquito killer are currently in progress steadily, strictly complying with international laser safety standards and electronic product specifications to ensure global market accessibility and user safety," it said in the update. Several of the backers on the comments thread below the post said they were concerned about compliance, citing the terms of service, which state that customers are responsible for ensuring it is legal to import the product into their own country. The project's creator, calling himself Jim Wong, responded directly in the thread with what AI text checkers determined to be an LLM-aided or generated message acknowledging the issue, but provided no assurances that Photon Matrix Lab would address it. "Thank you for the thoughtful message – and I hear your point. You're right that backers need clear certification information to make an informed decision, and we take that responsibility seriously." Wong went on to say that refunds were available until an order shipped, but made no commitment to verify the product's legality in each destination market, which to be fair, is not unusual, as many border agents will attest. In the past week, Indiegogo backers have also expressed concern that the company has not addressed whether the lasers could interfere with aircraft. Other backers claim repeated refund requests have gone unanswered, while several have questioned whether the product is legitimate. The Register contacted the company for more information. ®

Windows 11 has moved one step closer to matching Windows 10 after Microsoft slid a movable taskbar into the Release Preview Channel. The feature allows the Windows taskbar to be positioned at the top, left, or right of the screen, as well as at its default location along the bottom. Users have long requested its return after Microsoft dropped the capability from Windows 11. After floating around the Windows Insider Program for several months, the feature reached the Release Preview Channel late last week. Other taskbar tweaks in Release Preview include a compact option with smaller icons and reduced height, freeing up more of the desktop. Microsoft has not said when the enhancements will reach general availability, but their arrival in Release Preview suggests a wider rollout may be only weeks or months away. Microsoft has also added Start menu customization options, allowing users to choose between small and large layouts rather than accept the default size. The promised Windows Search updates are also present, making the service more "dependable" and allowing users to remove Microsoft Store and web suggestions from the results. The Release Preview also contains a couple of changes likely to please admins. Administrator protection lets users perform elevated tasks through just-in-time privileges. It is disabled by default but can be enabled through Intune or Group Policy. Microsoft has also pulled Windows Management Instrumentation Command-line (WMIC). WMIC's demise may prompt a sigh of relief, since miscreants have long used the utility to run commands and move around compromised systems. Deprecated years ago and already absent by default from new installations of Windows 11 24H2 and 25H2, WMIC has now disappeared from the Features on Demand list. Admins still using the utility may nevertheless feel the change. WMI itself is unaffected, and Microsoft recommends PowerShell as WMIC's replacement. ®

Despite the outcry over Capita's performance on the Civil Service Pension Scheme (CSPS), the UK government has awarded the beleaguered outsourcer a £31 million contract to build a public health contact center that could be called upon during another pandemic. The government estimates that a COVID-scale emergency could increase the center's total workload to £350 million. Capita announced its contract with the UK Health Security Agency (UKHSA) – an executive agency of the Department of Health and Social Care – to build and operate its Single Service Centre (SSC), an omnichannel contact center supporting the agency and other organizations. Due to begin in November, the deal has an initial three-year term. The award comes despite fierce criticism of Capita's performance running the CSPS. Its began administering the scheme in December last year, after which The Register exclusively revealed problems with its online systems. By March, it was clear the service was seriously failing, leaving some retired civil servants struggling to make ends meet. Capita won the seven-year, £239 million contract to oversee the CSPS in November 2023. The UK government has withheld £10 million in payments following the disastrous transfer of CSPS administration to Capita. It remains in dispute with the outsourcer over further payment reductions and has called in an independent auditor. In June, Capita missed a deadline to restore services to the required standard. The following month, it told MPs that service would return to normal for all but the most complex cases by September. UKHSA emerged from Public Health England in 2021 following criticism of its early handling of the COVID-19 crisis. UKHSA plans for the SSC to handle routine public health services while being able to scale up during future health emergencies and other national incidents. "The service will provide cross-government surge services during health emergencies and other national incidents across government," Capita's public statement said. An official procurement notice published in June said UKHSA had awarded Capita and US customer experience company Foundever places on a framework for the Single Service Centre. The maximum value assigned to each supplier was £157.5 million including tax, and the framework has an initial five-year term. During preliminary market engagement, UKHSA revealed more about the expected role of the SSC, which it said was "born from the Test and Trace service." Test and Trace was allocated a £37 billion budget over two years and relied heavily on consultancies including Deloitte. "In a pandemic, UKHSA must provide a service that communicates with the public on a person-to-person level, at a significant scale. Communication needs to be bilateral, citizens must be able to contact us for advice, order tests, or request therapeutics, such as Antivirals, and UKHSA will need to reach individuals for contact tracing or provide guidance," the notice said last year. "Should a health emergency occur on a similar scale to COVID-19, UKHSA has estimated that the value of the work would increase substantially from £39 million for the business-as-usual activities to an estimated £350 million, including tax." Capita said it would use a "centralized operating model" to handle customer contact, reporting and insight, workforce planning, and quality assurance. Its Capita Connect platform uses cloud infrastructure including Amazon Web Services and, the company claims, "advanced data capabilities to provide resilient, flexible and efficient operations." ®

The cost of agentic AI workflows is forecast to increase more than fivefold by the end of 2028 as users adopt more complex applications of the technology. As Nvidia and other tech giants push inference and agentic AI as the next stage of the AI wave, Gartner warns that the cost of implementing these systems will rise even as foundation models become cheaper. The analyst firm has cast its eye over the nascent world of AI agents – systems designed to act independently in pursuit of a goal – and sees multiple challenges ahead. Leaving aside the substantial security concerns, these software agents are considerably more complex than chatbots. Gartner believes falling model prices are tempting users to build more complex workflows, whose greater token consumption can outweigh those savings and drive up overall inference costs. In other words, tokens are becoming more cost-efficient, but those savings are not keeping pace with the rising cost of more advanced AI capabilities. The rate of innovation is outpacing the cost curve, Gartner claims. "The harsh economics of the inference paradox are exemplified by the differences between a simple chatbot and an AI agent," says Gartner senior director analyst Will Sommer. "Where a simple chatbot must read and interpret a query and quickly respond with a probabilistic reasonable answer, an AI agent must constantly reason, negotiate, and question itself," he explains. Those processes add up: routing a task to an agentic reasoning model increases inference costs at least fivefold, and potentially by much more as the task becomes more complex. Securing a return on investment from such advanced AI tools therefore demands either much greater returns than basic models provide or better optimization of inference, routing, and orchestration, Gartner warns. This could mean assigning each task to the most cost-efficient model capable of handling it. The move by some AI providers from flat-rate subscriptions to usage-based billing hasn't helped, as The Register reported last month. Token-heavy workflows can produce runaway costs under the new model. Perhaps it is no wonder Gartner predicted earlier this year that 40 percent of organizations would demote or decommission AI agents because of problems with the heavily hyped technology. The analyst biz also cheerily forecast last year that at least half of all generative AI projects would blow their budgets because of poor architectural choices and a lack of expertise, while most attempts to build custom models would be abandoned. ®

A cybercrook claims to have siphoned millions of employee records from the Microsoft Azure environments of major companies including McDonald's, Vodafone, Kyndryl, and Tata Consultancy Services. The alleged haul spans nine organizations and is being advertised for sale by a threat actor using the name "TheHatman," according to research published by Hudson Rock. McDonald's accounts for the largest alleged dataset on TheHatman's shopping list, with 1.7 million records purportedly up for grabs. Another 800,000 records supposedly come from Tata Consultancy Services, 425,000 from Vodafone, and 250,000 from HCL Technologies, with IHG Hotels & Resorts, Kyndryl, Gap, Hexaware Technologies, and Wyndham Hotels & Resorts rounding out the haul. Hudson Rock assessed the data as "highly likely authentic," citing corporate email addresses and structures consistent with exports from Microsoft Azure directory services. The records allegedly contain considerably more than names and work email addresses. Samples reviewed by the security shop reportedly include phone numbers, physical addresses, employee IDs, job titles, departments, office locations, reporting structures, group memberships, and service account details. Some records also reportedly identify accounts with Global Administrator privileges, potentially handing attackers a useful map of whom to target next. Even if the passwords aren't included, knowing who holds the keys to the kingdom makes for a handy phishing shortlist. How TheHatman allegedly obtained the information remains unclear. The attacker claims to have used compromised credentials, but Hudson Rock could not independently establish the initial access vector. It floated several possibilities, including credentials or session cookies stolen by infostealer malware, phishing, weak or absent multifactor authentication, and overly permissive third-party applications. Hudson Rock said its infostealer database contained compromised Microsoft cloud credentials associated with most of the named companies, although it could not link those credentials to TheHatman's alleged access. "Judging by the massive size of the organizations impacted, it appears highly likely that this campaign originates from targeted exploitation of Infostealer infections rather than a systemic zero-day vulnerability in Azure," said Hudson Rock. "If this were a widespread vulnerability, we would likely see a much broader spectrum of organizations impacted, including smaller businesses, rather than just these massive Fortune 500-level enterprises." The Register contacted all the organizations named by Hudson Rock to ask whether they were breached, whether the advertised data is authentic, and how any unauthorized access occurred. We've also asked Microsoft whether it is aware of a wider campaign targeting Azure or Entra customers. Tata Services sent The Register the statement it made to India's stock exchange [PDF] saying that the “Company has received threat-intelligence alerts alleging possible exposure of certain employee information." It added: The Company has investigated the matter and has not found any credible evidence of a breach of TCS systems or customer environments. The information referenced appears to be more than four years old and limited to basic employee information. There is no indication that customer data, customer systems, or TCS operational systems have been impacted. “The attacker claims to have used password spray and Multi-Factor Authentication (MFA) fatigue as the attack vector. The Company has had strong safeguards in place against such techniques for more than two years. Based on the current review, these controls remain effective, and the Company continues to monitor the environment closely." It said: “The Company will continue to assess any new information that becomes available and take appropriate action, if required. The Company remains committed to maintaining the security and resilience of its systems and to protecting the information entrusted to us.” TheHatman claims to have the data. How it might have walked out of nine corporate directories is the part nobody has explained yet. ®

Microsoft will retire Excel's COPILOT() function on September 14, barely a year after its preview debut. Introduced in August 2025, initially for Beta Channel users with a Microsoft 365 Copilot license and later for Excel for the web users through the Frontier program, the COPILOT() function let users send instructions to the company's AI assistant directly from a worksheet cell. Microsoft has now decided the Copilot side pane should be enough for anyone. Beginning September 14, 2026, the COPILOT() function will no longer be available. Microsoft had planned to make the function generally available in 2027. It has now updated the Microsoft 365 roadmap to say: "We have decided not to move forward with this feature. We apologize for the inconvenience." According to a Microsoft 365 Message Center post, "this change helps streamline the Copilot experience within Excel while continuing to provide AI assistance through a supported interface." "Customers can continue using Copilot in Excel through the Copilot side pane, which provides many of the same AI-powered capabilities, including summarizing text, classifying data, generating content, and retrieving information from the web." Since the function remained in preview, spreadsheet wranglers should not have relied on it in production workflows. Anyone who did now has some formulas to replace. In its original announcement, Microsoft wrote: "Its output should be reviewed and validated for accuracy, especially for critical business decisions or reports." Hardly music to the ears of spreadsheet users, for whom accuracy tends to matter, but par for the course with AI services. Microsoft has spent recent months trying to make Copilot behave more consistently across its productivity applications. It has also responded to some user feedback – the Dynamic Action Button can, for example, be banished to the toolbar rather than left floating atop the content. With similar capabilities remaining in the side pane, Microsoft evidently decided that a dedicated worksheet function was surplus to requirements. Google Sheets still offers a similar AI function. The search giant announced the feature in June 2025. ®