The Register

OPINION On June 24, 2021, Microsoft announced Windows 11, unveiling a new and controversial operating system. Five years on, how has that worked out for you, Redmond? Windows 11 has always been a problem child for Microsoft. It was announced in June 2021 and became generally available on October 5 that year, while much of its customer base was still reeling from the COVID-19 pandemic. At the time, The Register called it pointless rather than a point release of Microsoft's flagship operating system. Why? Because Windows 10 was more than adequate. Microsoft's apology for the Windows 8.x era was… fine. It mostly worked without difficulty. It lacked the user-experience missteps of its predecessors and was an architectural step up from Windows 7. And, most importantly, the operating system didn't trip up a user's workflow. There is an old adage: "If it ain't broke, don't fix it," but Microsoft set to work fixing Windows 10 regardless, and the result was Windows 11. The user experience has irked users ever since. Did you like being able to move the taskbar around in Windows 10? Tough – in Windows 11, you'll have to learn to love where Microsoft stuck it. How about the Start Menu? Again, Microsoft knew best and redesigned it. In the last year, Microsoft appears to have realized that its actions have alienated users and promised to restore eliminated user interface elements, such as the movable taskbar. It hasn't, however, gone back on another Windows 11 feature – the infamous hardware requirements. While Windows 11 contained plenty of software elements to annoy users, it was the company's decision, on security grounds, to render hardware perfectly capable of running the operating system obsolete at a stroke that really angered users. Even hardware (including some of the company's own) that was still on sale at the time wouldn't work. The company demanded TPM 2.0 and warned that anything older than an eighth generation Intel CPU (or equivalent) would not make the cut. Then and now, the decision carries an arbitrary air, particularly as several workarounds emerged, revealing the requirements to be the technically unnecessary decisions they were. More than anything, Microsoft's hardware requirements slowed the operating system's adoption, as hardware that ran Windows 10 perfectly well was rendered obsolete overnight. In the end, it took until 2025 for Windows 11 to overtake its predecessor in market share, and until 2026 for the gap to widen. Much of the change in market share is likely due to hardware replacement cycles and the end of mainstream support for Windows 10 on October 14, 2025. Microsoft's operating systems have followed a set pattern over the years. Windows XP was good, Windows Vista was not. Windows 7 was good, Windows 8.x was not. Windows 10 – good. Windows 11 – oh dear, it seems it was always destined to be a bit of a duffer, even without Microsoft loading it with ads and AI. While hindsight has made Windows 10 seem rather good, retrospectives are unlikely to be so kind to Windows 11, which marked an era in which Microsoft took its eye off the desktop to focus on shinier, AI-related things. Microsoft has already dropped Copilot branding from products like Notepad, an acknowledgment that the assistant is not the welcome pal in every place it is forced into. The same could be said for Windows 11, which has become a byword for iffy quality and bad management decisions. Based on the last five years of Windows 11, Windows 12 should be a beacon of light. Right? ®

The next major release of Deno, a JavaScript/TypeScript runtime, will include new commands to build cross-platform desktop applications using web technology. Deno desktop will compile an application from a plain TypeScript file or from frameworks such as Next.js, Astro, Deno Fresh, TanStack Start or Vite SSR (server-side rendering). A distinctive feature is that the native WebView is used by default, rather than bundling the Chromium Embedded Framework (CEF). The advantage is much smaller applications. In our quick test, a compiled application using WebView on macOS was around 68.5MB, whereas using CEF, which is also an option, the size was 308.9MB (though the docs suggest 150MB as typical) and startup was slower. The trade-off is that CEF provides all the features of Chromium and consistent rendering across platforms. Without CEF, it is a challenge to ensure the WebView on all supported platforms will work as expected, particularly since users may install on older machines with out-of-date native web browsers. This is a particular problem with macOS and Safari. Deno desktop has a further option, called Raw, which has no web engine. Developers get window management and can draw a user interface with WebGPU, the Skia library, or their own custom rendering. A Deno desktop application, other than one using Raw, includes a local web server, which means that developers can easily port web applications. According to the docs, the cost of "a single network hop" is negligible in most cases. That said, there is also an option for in-process binding, which exposes a Deno function directly to the WebView or CEF. Deno Desktop supports native menus and context menus (right-click), native dialogs for alerts and confirmations, and notifications via the Web Notifications API, which appear as native notifications. File picker dialogs are not yet supported, other than by the web file input type, or drag and drop. There is no separate clipboard API yet. Mobile applications are not yet supported, though it is planned. There are plenty of existing options for building desktop applications with web technology, including the popular Electron used by many well-known applications but sometimes disliked for its high resource usage. The Deno desktop docs include a comparison with Electron, Electrobun, Tauri, and Dioxus. Deno desktop looks well thought-out, though it is not yet stable, and we noticed issues such as the window close button not working in macOS using WebView, as well as friction trying to get it working with some web frameworks. The broader question, perhaps, is whether this will help the Deno project gain market share. It began as an effort to improve on Node.js, and has been somewhat upstaged by the Bun project now owned by Anthropic. One of the problems for Deno is that the Node.js ecosystem is so entrenched that the project has had to retrofit Node.js compatibility, diverting engineering time from other potential features. One Deno user, Hong Minhee, posted that they liked Deno initially because it avoided the pain of Node.js. "No configuration files, no node_modules, no agonizing over which package manager to use." Now though, "Deno is spending more and more of its energy catching up to Node.js," while on the other side of the camp, Node.js has added features such as TypeScript compatibility. "Why didn't Deno hold the course?" they ask. The quick answer is that the pull of the established Node.js ecosystem was too strong. Deno desktop looks promising and could provide another reason to use Deno if it works well, though there is also a risk of further diluting the resource available for the core runtime.®

Prestigious journal Nature has published a peer-reviewed critique of Microsoft's claims to have made quantum computing breakthroughs – and the scientist who wrote the paper has essentially said Redmond got it wrong. Microsoft made its claims of a quantum breakthrough in February 2025 when it revealed tech called Majorana and predicted "this breakthrough will allow us to create a truly meaningful quantum computer not in decades, as some have predicted, but in years." The software giant's approach to quantum computing involves Majorana particles, subatomic particles that scientists have not observed directly. The company has pursued this approach for years, but experienced reversals that led to the retraction of some papers. Last year, however, Microsoft claimed it had both observed Majorana particles and harnessed them in a quantum computer. Criticism of that claim was swift and sharp: we reported boffins willing to go on the record as describing Microsoft's work as "unreliable" and perhaps even "fraudulent." Microsoft insisted its work is sound and in early June 2026 announced Majorana 2, a "next-generation topological quantum chip" it developed with the help of its own agentic AI. The Windows giant revealed that work after being given a right to reply to a critique of its 2025 Majorana announcement by Dr Henry Legg, a lecturer at the University of St Andrews. Nature accepted Legg's paper on April 20 and scheduled it for publication on June 24. Titled "On the robustness of topological gap detection via transport," Legg's analysis suggests Microsoft got it wrong. "Last year they claimed to be years, not decades from a 'topological quantum supercomputer,'" Legg told The Register in an email. "My feeling is that they are centuries, not decades away. If it works at all – and, based on what I have seen, the most likely scenario is that it doesn't work." Based on his analysis of the research Microsoft published in 2025, Legg argues that the company's claims about finding and being able to control the elusive Majorana particle to build a topological superconductor do not withstand scrutiny. "I demonstrate that Microsoft's tune-up software is flawed and that coding errors resulted in incorrect statements to peer reviewers," said Legg. "Raw data, which was omitted from the original paper, also appears to indicate Microsoft's devices contain considerable disorder and are not compatible with the existence of a topological gap. In other words, the prerequisites for Microsoft's claims do not appear to be met, but this was obscured because this data did not appear in the original publication." Essentially, Microsoft has proposed a Topological Gap Protocol (TGP) that can be used to detect the phase transition deemed to be a prerequisite for conducting quantum calculations using Majorana particles. Legg argues that based on his analysis of underlying transport data (measurements of particle change) – omitted from the original publication – Microsoft chose to focus on results that supported its thesis and ignored data that could be interpreted as a negative result. As he notes in his critique: "The TGP plotting code was set to highlight only the largest purportedly topological region." "The primary consequence was the omission of other regions that passed their tune-up protocol (the TGP)," said Legg. "When peer reviewers asked if other regions existed, Microsoft inaccurately stated that they had investigated the only region passing the protocol within the explored range. This was not correct." Legg also argues that Microsoft mishandled its code. "The code antisymmetrized bias voltage based on array index rather than physical value," his analysis says. In other words, Microsoft's researchers made a basic programming mistake by evaluating the array index – the number identifying a value's position in an array – instead of the value to which the index refers. "There were two pretty basic Python programming errors that hid these alternative regions," Legg explained. "Their plotting software was hardcoded with a filter (zbp_cluster_numbers=[1]) that forced it to display only the single largest region, concealing other successful results from their phase maps. Changing this to zbp_cluster_numbers=[1,2] shows already a second region." Legg added: "The TGP software transformed the data by simply reversing a Python array (x[::-1]) based on its index position, ignoring the actual physical bias voltages." In a statement provided to The Register, Dr Chetan Nayak, technical fellow and corporate vice president of Microsoft's quantum hardware group, said: "We stand by our results and our roadmap." "At the end of the day, success is the delivery of a scalable quantum computer. We are confident in our ability to execute against our roadmap and proud of our continued engagement with DARPA, which moved Microsoft into the final phase of its Quantum Benchmarking Initiative after independently evaluating our results – those in the public realm and proprietary – with a team of highly qualified experts. Skepticism and rigor are hallmarks of the scientific process, which we appreciate and have supported from various academics. We have participated in dialogue and our thorough rebuttal was accepted and published by Nature." Microsoft's rebuttal disputes the validity of Legg's analysis. The software colossus argues its signal measurements were not intended to be exhaustive and that the "minor off-by-one-pixel bug in our TGP processing" is inconsequential. The response concludes: "In summary, Legg centers on a selective examination of transport tune-up procedures and narrow interpretations of isolated phrases in our referee correspondence, rather than the physical mechanisms underlying the experiment. It relies on unsubstantiated claims about our transport spectra while not engaging with the capacitance measurements at the core of our study, and its alternative treatment of the transport data is inconsistent with more rigorous analyses of the same datasets. Critically, Legg offers no alternative physical model capable of reproducing the capacitance signal or the RTS phenomenology, and does not constitute a substantial scientific challenge to our findings." Legg thinks that criticism is unfounded. "They attempt to dismiss these issues as minor bugs, and retrospectively adjust their evidence hierarchy," he said. "In short, Microsoft's reply essentially argues that because they observed a specific capacitance measurement, the prerequisites to do so must have been met. I hope, despite the complexity of the topic, their circular reasoning is clear." The announcement of Majorana 2 has not changed Legg's assessment of Microsoft's work. "Majorana 2 is not available to customers and it is not proven to even be a single qubit," Legg said. "Their preprint, which should not really be given any credence given that it is based on a single device, does not even claim an X-measurement (which they did eventually for Majorana 1 last year, but that preprint has also not yet been published). Essentially, their claim of '1,000 times more reliable' refers to the lifetime of a classical bit (the parity of the state). There is no evidence this is a qubit and can hold a superposition. The classical bits in my computer have very long lifetimes (years!), but it does not make them good qubits." "For Majorana 2, one has to ask why they do not report the X-measurement, since Microsoft were obviously aware it was so important for their claims last year. I think it's very reasonable to assume that they did attempt the same supposed X-measurement with their Majorana 2 device and it didn't work out. That's not surprising because, based on everything I have seen, it all looks like disorder physics and they have not shown any kind of control over even a single qubit." ®

AI models used to help diagnose medical conditions have a problem: They’re ready and willing to identify patients whose data was used to train them. German researchers reported in a Nature paper published Wednesday that discriminative AI models - those used to classify data and make predictions about new inputs based on their training sets - are particularly susceptible to membership inference attacks (MIAs) that query the models in an attempt to figure out whether a particular datapoint is included in their training sets. What that means for medical AI models is that any patient whose data is used to educate the bot could be exposed, leading to details about their medical history and diagnoses being leaked. In an analysis of seven medical AI datasets consisting of images, ECG records, and general electronic health records, the team determined that individual patients targeted by such attacks can be identified with “near-perfect attack success,” which they explain flies in the face of how such models are evaluated for safety. “The fact that MIAs can achieve near-perfect success rates for individual patients is not adequately captured by the standard evaluation protocol, which measures attack success in aggregate across records,” the researchers said. Based on their findings, they conclude, reporting standards for AI privacy audits need to change. It gets worse, too: Patients in the dataset are generally easy to identify and, unsurprisingly, those underrepresented in medical AI training data are even easier to finger than those whose data doesn’t stand out. Underrepresented groups can include those in a number of sensitive categories: Race, insurance status, sex, the protocol used to conduct medical imaging, and certain disease statuses can all function as outliers that make it easier to identify individuals. “Generally speaking, privacy risks from MIAs become more severe as a model’s training cohort becomes more specific,” Technical University of Munich AI in Healthcare and Medicine chair and paper lead author Moritz Knolle told The Register in an email conversation. “You could imagine … scenarios where membership in a training dataset reveals that someone has a dormant genetic condition such as Huntington's disease, depression, or attended a specific, specialised treatment clinic.” In other words, exposing healthcare AI training data could be used to identify those with sensitive health conditions, spill secrets they may not want public, or otherwise fuel discrimination. To make things even worse again, the larger the dataset, the easier it is to expose records, and “the magnitude of this change in patient-level risk was previously unknown” in larger models. The privacy devil in the data details This is bad and all, but it’s not necessarily the end of the world, as performing an MIA attack on a medical AI model supposes the attacker already has a few things at their disposal, namely at least some medical data belonging to the people they want to identify. “To conduct a MIA an attacker needs access to a target data point,” Knolle confirmed to us while also noting that their paper revealed access to a full patient data point isn’t needed, in contrast to what was previously believed. “In our paper we show that an attacker with partial access can still successafully conduct MIAs.” The MIA attack itself, as detailed in the paper, relies on medical AIs being more certain of their predictions if the input data is already part of their training set. A potential attacker, then, simply peppers an AI model with obtained patient data, checks the confidence level, and surmises that said patient is part of the training data. “An attacker conducting a MIA does not need to know who the data belongs to that they are trying to conduct the MIA with,” Knolle explained. “In fact, all the dataset we use in our study were anonymized.” Anonymized in the datasets, but not the target data, that is. As explained in the paper their MIA attacks were largely error-free at the individual patient level, meaning confidence levels are an accurate way to figure out if a particular patient's data is part of a training set. “The attacker would simply need access to someone’s blood test results, or part of these results” in order to infer inclusion, Knolle said. Of course, they have to get that data first, but given how frequently healthcare data is exposed in breaches, it’s not exactly hard to imagine a bad actor getting ahold of something they can use. “Given that medical data is not always securely stored it is not unthinkable that an attacker could get access, for example, by gaining unauthorized access to the database of your general practitioner after they performed a routine blood test,” Knolle said. How to protect patient data? Asked what he hopes this research accomplishes, Knolle told us he just wants the medical world to understand that AI training data needs to be better secured. “I hope that the medical AI community will start to take privacy risks seriously and that risk mitigation techniques are used in situations where they are necessary,” Knolle said. The researchers make several recommendations for how to do this, like through the use of differential privacy frameworks that are designed to mathematically guarantee training data remains anonymous - a key consideration if medical AI firms want patients to trust them with their data. As mentioned above, the team also wants to see privacy audit standards change to consider individual-level data, not just aggregate privacy risks. Alternatively, medical AI training data could just be compiled so that underrepresented groups are better represented, Knolle said. “There are many situations where a successful MIA represents a small or negligible privacy violation,” Knolle noted. “These are situations where AI models are trained on large, general populations in which both healthy and diseased individuals are represented in sufficient numbers.” Representation, in other words, definitely matters when it comes to keeping patient data private. ®

The Home Office is looking for a software engineering boss ready to support systems that when they go down can ruin the day of many thousands of people and make headlines internationally. The Director of Engineering for Home Office Digital will be responsible for managing and deploying nearly 1,000 engineers working on more than 600 systems. Among many other things, these systems enable 76 million e-gate border crossings, 140 million police checks on people, vehicles, and property and seven million passport applications every year. Since February, the UK has required many foreign nationals to apply for electronic travel authorisations, which as well as blocking some dual citizens from getting back into Britain adds another item to the list of what the job application pack calls “some of the UK’s most critical and well-used public services.” The pack adds that the department is looking for “an experienced engineering lead with a deep understanding of how to operate in a fast paced, security critical and highly visible environment” adding that “this is a leadership role of exceptional criticality and complexity”. It involves managing a £60M annual budget and a software engineering function currently known as SAS, which may or may not be a naming coincidence. The successful applicant will be paid £120,000-£150,000 a year plus a 29 percent employer pension contribution, and can be based in London, Croydon, Manchester or Sheffield. He or she will be expected to serve for at least three years, although this is not a contractual requirement, and must be British or from a European or Commonwealth country with the right to work in the UK. Applicants should email a CV and 1,250-word statement of suitability to the Home Office’s recruitment firm Global Resourcing by 23.55 BST on Tuesday 21 July. You can take a look here. ®

The British government wants "trusted" news sources to be made more prominent on social media in plans that seem set to cause controversy with free speech advocates. In a Green Paper published Tuesday, the Department for Culture, Media and Sport (DCMS) laid out its aim to improve access to reliable news sources on online platforms. This is part of broader reforms to the UK's public service media system, intended both to ensure it continues to serve the British public and to support an eventual managed transition to internet-based TV. DCMS observes in the report "Watch this Space: A new strategic direction for UK media" that there has been a shift in how people consume news. For 75 percent of young people aged 16-24, social media is their main source, while more than half of UK adults now include social media as one of the ways they get updates. As Reg readers are well aware, while social networks provide access to a greater range of news sources, they are also a superb conduit for misinformation and disinformation, with "the potential for less accurate material to replace trustworthy sources as content is increasingly shaped by algorithms and AI." Because of this, the government is inviting feedback on its proposals. These would require social media companies and video-sharing platforms to make sure that news content from public service media (PSM), which includes the BBC, ITV, STV, Channel 4, S4C, and Channel 5, plus other trustworthy providers, is "prominent and easy to find on their platforms." How this would be achieved is up for discussion. DCMS suggests that it could involve a range of national and local news publishers being given prominent placement in search results or recommendation feeds when people search for news. The government claims to be keeping an open mind on this, seeking views on whether it can be achieved voluntarily, whether all PSM content should be prominent, and which platforms ought to be in scope. The Green Paper says that PSM providers are held to higher regulatory standards than other content providers, and this in turn increases their costs. Historically, they were also more prominent to audiences, but "in a platform environment driven by algorithmic recommendations, media is frequently optimized for engagement rather than public value." The paper claims that the risk is stark, especially for young people, and that there is broad societal consensus that such audiences should have guaranteed access to "high-quality, safe and educational content." This is consistent with the government's recent actions to protect children online, it further claims. But critics will argue that the government is trying to control the narrative, forcing social media platforms to give higher prominence to news sources that may be seen as part of the establishment or less likely to be critical of the government itself. We can only imagine the reaction of X owner and "free speech absolutist" Elon Musk, for example, who has openly clashed with the UK government on numerous occasions. We asked X for comment on this article. Jemimah Steinfeld, chief exec of campaign group Index on Censorship, said the government proposals are a cause for concern. "While the proposal sounds well and good – we all want less mis and disinformation online – who gets to decide and define what is 'trustworthy' is ripe for abuse," she told The Register. "Just look elsewhere for evidence. In the USA, for example, Donald Trump calls CNN and other heritage news companies 'fake news.' Do we essentially want to give leaders more power to control what we do and don't see?" Social media news reforms must therefore be "transparent and proportionate," according to Hanna Basha, a partner at London law firm Payne Hicks Beach. "The government is right to consider how trusted and regulated news providers can remain visible and accessible to the public," she said. "However, any new prominence requirements will need to be carefully balanced against freedom of expression and existing obligations on platforms under the UK's online safety framework, while providing clarity about how trusted news sources are identified and promoted." Also part of the Green Paper consultation is the future of television. DCMS notes the TV landscape is undergoing "a profound shift" from traditional broadcast to online linear viewing, creating a hybrid landscape between IPTV, traditional digital terrestrial television (DTT), satellite broadcasting, and cable. While online-only access is growing, a significant number of viewers continue to rely on traditional broadcast platforms. But maintaining the DTT network beyond a certain point means broadcasters carry the costs of multiple distribution methods to an ever-shrinking number of households relying on DTT. It believes this point may come as early as 2034, but that the government will set out the plan for a managed switch-off of DTT services, either on expiry of current licenses on December 31, 2034, or following a time-limited extension to December 31, 2044. The Media Green Paper public consultation started on June 23 and will run for ten weeks until midnight on August 31. Those wishing to respond can do so here, or by sending an email to watchthisspace@dcms.gov.uk. ®

Gartner has slammed AI vendors' lack of transparency, saying developers are facing sharply increased costs from coding agents. Since the main AI coding agent vendors have shifted from seat-based licensing to consumption-based pricing, developer teams now face highly variable cost structures. Developer teams face an emerging problem of escalating costs, said Nitish Tyagi, senior principal analyst at Gartner. AI coding bills were leaping from $20 or $100 to $2,000 to $5,000 per developer per month, while in extreme cases, the bill might hit $20,000 in token charges. However, software engineering departments get little insight into how token consumption is calculated and billed, making it difficult for them to forecast and control costs accurately. AI coding vendors have not yet delivered built-in features to allow developers to optimize costs in AI coding agents, resulting in cost escalation, Gartner said. "None of the vendors have incredible features when it comes to cost optimization," said Tyagi. Instead, he said vendors were focusing on the concept of "tokenmaxxing" to "boost the high" of token consumption, suggesting that if developers increase the number of tokens, they will increase productivity gains. "There is no direct relation between the increase in token consumption and an increase in productivity gains," he said. Gartner recommends developer teams optimize token consumption and adopt strategies such as context engineering practices, where software engineers improve the input context provided to AI systems. Another recommended strategy is model routing, where engineering and platform teams direct simpler, high-frequency tasks to smaller models, using frontier models only for complex, high-value work "All of these things will improve the output quality, and, therefore, will increase the productivity gains as well, so while there is no direct relation between tokenmaxxing and productivity gains, there is a relation between optimization of token consumption with the output quality," Tyagi said. As a result of the lack of cost optimization tools among vendors and consumption-reduction strategies among users, AI development is in a situation where a developer's coding agents may cost more than they earn, at least in some parts of the world. Gartner predicts that by 2028, AI coding costs will overtake the average developer's salary due to rising LLM token consumption and the shift to consumption-based licensing models. "We're not saying AI token cost will be higher than every developer's salary on the planet, because US salaries tend to be higher than in India, for example. But current token costs are already more than most of the salaries in India," he said. However, the cost of coding agents does not vary according to where they are consumed, meaning AI coding costs for a developer in India may be equivalent to the salary of an engineer with four to six years' experience, Tyagi said. "We're not saying AI token costs will be higher than every developer’s salary on the planet, because US salaries tend to be higher than in India, for example," Tyagi said, adding that token costs do not vary by location, and in India they may already be equivalent to the salary of an engineer with four to six years' experience. ®

The Metropolitan Police Service (MPS) will start using static live facial recognition (LFR) cameras in London's West End and Soho by the end of this year following a six-month pilot in the south London borough of Croydon. Static LFR involves the police temporarily attaching cameras to lampposts or similar infrastructure, with the feeds monitored remotely and officers on the ground stopping people whom the technology matches to images on its watchlist. The MPS said that each of the 24 deployments in central Croydon between October 2025 and March 2026 used a bespoke watchlist created up to 24 hours in advance and deleted afterward. Civil liberties campaign group Big Brother Watch, which in April lost a High Court challenge to police use of LFR, said the force was rushing ahead with deployment before Parliament has passed legislation regulating the technology's use. "We are calling on the Met to stop this experiment until, at least, Parliament has spoken," Jack Coulson, the group's head of advocacy, said in a press release. "Policing by consent is a cultural inheritance we must protect. Permanent biometric surveillance of the public square is incompatible with that ideal." He highlighted the case of Alvi Choudhury, a Southampton man arrested and held for ten hours in January after a retrospective LFR system run by Thames Valley Police matched him to a crime committed in Milton Keynes, a city he had never visited. "It is predictable, given the technology's racial bias, that Mr Choudhury was confused for another Asian man," said Coulson. The MPS said that in Croydon more than 470,000 people walked past the LFR cameras, leading to 173 arrests and one false alert, which resulted in officers stopping someone without arresting them, realizing the mistake, and letting them go. The force added that one of those arrested, a registered sex offender who was communicating with a child under 16, was subsequently sentenced to two years in prison in May for breaching a sexual harm prevention order and making indecent images of children. MPS Commissioner Mark Rowley said on June 24 that the force planned to "significantly step up our use of technology to fundamentally change how we protect the public" through the use of live LFR, a city-wide emergency services drone network, and AI to analyze the footage from the capital's one million CCTV cameras. Rowley added that the force needs to spend more on technology but its budgets for doing so have been repeatedly cut, with spending of around £6,000 per person compared with budgets of more than double that at some government agencies. Earlier this month, the commissioner said the MPS would have to cut around 700 frontline posts after London's deputy mayor for policing and crime, Kaya Comer-Schwartz, refused to approve its plan to award a major contract to controversial US supplier Palantir. ®

The UK Competition and Markets Authority's (CMA) investigation into Microsoft's business software ecosystem has attracted a full range of comments from the unsurprising to the hyperlocal. The watchdog is looking at whether product bundling, the alleged embedding of in-house AI products like CoPilot, and licensing practices are limiting UK customer choice. Asking for submissions from customers and rivals allows the regulator to weigh Microsoft's arguments against the concerns and grievances of those customers and rivals. The regulator has now published those comments. Google reckons [PDF] Microsoft is "effectively a gatekeeper and uses its entrenched position to steer captive users toward its own cloud and artificial intelligence (AI) solutions," while Killinghall Parish Council complained [PDF] about an unplanned annual cost of £1,100 due to the purchase of additional Microsoft services "to use Microsoft Teams effectively." The submission stated: "We are concerned that Microsoft's refusal to integrate fully with third-party products creates unnecessary additional costs for Parish Councils." Other comments came from the Browser Choice Alliance [PDF], which stated: "Microsoft's tactics leverage its dominant Windows OS and M365 productivity software suite to frustrate users' choice of browsers." Mozilla complained about Windows 11's habit of defaulting to Edge, particularly during the upgrade flow, and Collabora called on the CMA to "explicitly include both APIs and document format standards (Office Open XML or OOXML) within the scope of the Productivity Software Suite description." "Because Microsoft's own applications handle OOXML formats as de facto standards with undocumented complexities, a significant technical barrier is created for third-party suites trying to interoperate seamlessly." And so it went on, with other businesses such as messaging service Element and browser vendor Vivaldi weighing in, and groups including Open Web Advocacy and the Open Cloud Coalition (previously described by Microsoft as a Google lobby group) having their say. For its part, Microsoft stated [PDF] that it "faces substantial competitive constraints across all five digital activities." In productivity software, it said the likes of Google Workspace and LibreOffice compete "vigorously." Windows faces competition from macOS and Linux, and Linux is a far more widely used server operating system. Databases, such as PostgreSQL and MySQL, compete with SQL Server, and Okta makes a decent fist of Identity and Access Management. Furthermore, "Microsoft's commitment to openness and interoperability is real, deep and growing," and, significantly, the company stated, "AI is making competition more intense, not less." The CMA's investigation could result in a Strategic Market Status (SMS) designation for Microsoft, which would allow the CMA to implement interventions to support competition. The wide-ranging investigation kicked off in May and is looking into Microsoft's provision of business software to UK organizations, as well as bundling, defaults, and interoperability. The CMA stated that it was important for customers to be able to "mix and match software and AI services from a broad range of competing suppliers." And then there's licensing. An SMS designation would allow the CMA to intervene on a concern arising from its cloud market investigation – Microsoft's alleged use of software licensing to reduce cloud competition. The process has a few more months to go. The proposed decision will be published for consultation in October, and the final decision report (and any SMS decision notice) will be published in February 2027. ®

Irresistible force, meet immovable object. Tech leaders are under pressure to satisfy growing demand for AI while keeping a lid on costs. That is becoming harder as Anthropic, OpenAI, and GitHub shift some services away from flat-rate subscriptions toward usage-based billing. Database vendors claim they can help by cutting the number of calls made to AI models and handling the new workloads generated by developer agents. According to IDC research director Devin Pratt, "demand for the underlying capability is strong, because agentic adoption is already broad." Around 79 percent of organizations are either investing significantly in agentic AI with a set budget or already running agentic applications in production, according to IDC. "The appetite for the data infrastructure beneath it is real," Pratt said. "The open question is whether the specialists win or the capability gets absorbed into the platforms enterprises already run, much as it did with vector databases." Among the specialists taking on the omnipresent cloud platforms and their embedded data services is Pinecone. It carved out a niche as a vector database vendor for users wanting to put LLMs into production. Although that term became commonplace among mainstream database companies, Pinecone always argued it maintained a technical advantage for the biggest use cases. Building on top of its vector database technology, the vendor has recently launched Nexus – a "knowledge engine, not a retrieval system" – and embedded it in Microsoft's OneLake, a hybrid data lake and data warehouse environment. Pinecone's idea is that by compiling a knowledge base of an organization's data structure and content, its technology can avoid burning through tokens back and forth between the data and AI agents. Nexus is designed to structure, contextualize, and compose specialized contexts – derived artifacts – in advance of agent demand. Pinecone product veep Jeff Zhu told The Register the idea was to prevent agents from repeating the same work to understand the structure of business data and its context. "All these coding agents, for example, are really good at doing a bunch of exploratory work if you ask them a question," Zhu said. "It's going to make a call, get the table schema, do some exploratory work, figure out what the top rows of this one table are, and ultimately it will eventually get to the right answer most of the time, but it's going to burn through a bunch of tokens, because every single time it creates a specific answer to a question, it has to understand your business context and rediscover it every single time." Nexus provides a semantic layer of business data for a given use case or outcome. Imagine a finance analyst agent versus an HR agent: it is very possible they could use the same data, but they would want very different outcomes as a result, Zhu said. "Our system is an engine that can support any number of use cases. Based on your particular use case, and the tasks that you want to accomplish, we use your data sources – which can be SQL databases, unstructured documents, PDFs, and so on – we build a task-specific context which is used for that individual agent's job or role." IDC's Pratt said that the hard part of agentic deployments has shifted from the model to the data plumbing around it. Agents reason continuously and act on live data, and the traditional split between operational stores, analytical warehouses, vector indexes, and the pipelines stitching them together was built for humans, not software running in loops. A recent IDC Data Management survey found that the two biggest data roadblocks IT leaders name for scaling generative and agentic AI are security and compliance constraints and cost. Fragmentation compounds it, with nearly two-thirds of organizations running 11 or more distinct database technologies. Pratt said Pinecone is betting that retrieval as practiced today does not scale to agents and that, rather than have an agent rediscover and re-read raw content on every call, Nexus does the reasoning once, upstream, and stores reusable, task-specific context. The idea is credible, and the positioning is smart, moving Pinecone from a vector database to a knowledge layer. Its query language, KnowQL, also includes a budget primitive, and the platform gives a single dashboard for token usage and spend. By doing that work once rather than on every inference call, the design directly addresses one of IT leaders' chief concerns: cost. "What is interesting about Nexus is that it treats cost as a design constraint, not an afterthought, with a budget control and token accounting built into the query layer. That is exactly where IT leaders feel the pain," Pratt said. Changes in the wider tech landscape have previously forced new ways of looking at databases and related software. The market responded to the so-called Big Data problem – the staggering volume of information generated by mobile devices and web clickstream data – with Hadoop and Apache Spark. The former has largely been superseded; the latter is still very much with us. The demands consumers and businesses place on globally scaled web systems prompted a generation of NoSQL databases like Couchbase, Cassandra, and MongoDB, as well as new distributed relational databases. Another database vendor hoping to help tech teams manage the escalating costs of deploying AI agents is Tiger Data, the company behind the PostgreSQL time-series database TimescaleDB. It has built Ghost, a technology designed specifically for developers working with AI agents. The company argues that agents helping build software experiment constantly and need isolation to do it safely. When an agent fails, the blast radius should be one database, not a shared environment that other agents and humans depend on. The Ghost platform offers instant PostgreSQL databases with fast forking, which agents can access through the Ghost CLI or MCP server. It comes with a terabyte of free storage. Ajay Kulkarni, co-founder and CEO of Tiger Data, told The Register the company had also adopted a new database charging model better suited to AI agents, which typically produce spikes and falls in demand. "Instead of being packaged at the traditional database level, we provide usage-based pricing at the compute-hour level, so no matter how many databases you have – it could be one, it could be 50 – it'll cost the same, and you'll just be metered by how many compute hours you consume," Kulkarni said. Tiger Data offers a free tier with 100 compute-hours per month. Users can pay for additional usage in 15-minute active windows. "That really allows for experimentation and allows for this idea of multiple databases, with every agent getting its own database, because you're not paying on that dimension, you're paying on the compute-hour dimension," he said. IDC's Pratt said Ghost solves a different but equally concrete problem. It gives each agent or task its own disposable PostgreSQL database, allowing an agent to branch a dataset in seconds and discard it afterward. Because billing is based on usage rather than the number of databases, Tiger Data offers a practical way for organizations to stay with PostgreSQL. However, many Postgres vendors are racing toward the same capability. "Ghost is the most pragmatic of these. It hands each agent a throwaway PostgreSQL workspace to experiment in without risking production, and by staying on PostgreSQL it asks IT teams to learn almost nothing new," he said. Pratt argued that agent-ready data infrastructure is turning into its own category, and the platform vendors are the ones to watch. On the context and vector side, Pinecone competes with companies such as Weaviate and Qdrant, as well as the pgvector ecosystem inside every Postgres distribution. On the agent workspace side, Tiger Data's Ghost sits alongside Neon, now owned by Databricks, and Supabase. Bigger platform vendors such as Snowflake, Oracle, and Microsoft are also absorbing these capabilities into the stacks customers know well. "The independents are defining these categories, but watch the platform vendors. Most organizations tell us they expect to do their vector work inside the database or lakehouse they already run, not in a separate specialist tool," Pratt said. Aaron Rosenbaum, Gartner senior director analyst, said that building this context layer is critical for both token efficiency and better answers. "We've seen significant development across the industry; Snowflake introduced Horizon Context; Databricks introduced Genie Ontology, and we will see significant growth in the quality and depth of new tools introduced throughout the year. While most of the leading data platform vendors have their own solution, there will be enterprises that look for solutions independent of those platforms that Pinecone Nexus can address," he said. He also noted that there was a string of vendors examining the different DBMS workloads created by agents compared with those created by humans. "It's driving new innovation across the industry right now," Rosenbaum said. As technology leaders try to negotiate better deals with AI providers to keep costs down, they can also look to their data platforms to ease the financial burden created by the often-unmanaged surge in demand from AI workloads. "There are many options available today from all the major providers, and no clear leader has emerged. History has shown that the established DBMS vendors have been very agile at supporting new workloads, and it's likely that will continue," Rosenbaum said. ®

Alpine Linux 3.24 is out, bringing a new desktop environment that should make for a very high-performance combination. Version 3.24.1 followed just four days later to fix some OpenSSL security issues. The 3.24 series contains new versions of GRUB, LLVM, Rust, Go, Qt, and NGINX. Alpine isn't an entirely GNU-free Linux distro, but it doesn't use many components from the GNU Project: it's systemd-free and doesn't use the GNU standard C library, replacing it with musl libc. Even so, version 3.24 offers the latest GNOME 50 desktop, as well as version 6.6.5 of KDE Plasma. New for this release is "Epoch 1" of System76's COSMIC desktop. The Reg FOSS desk took a look at the new environment at the end of last year and came away impressed. COSMIC is the first pure-Wayland desktop environment we've tried that we would be willing to use full-time. We've tried it on some quite old machines, including a ThinkPad X220 with an Intel GPU and a ThinkPad T420 with an Nvidia GPU, and while we have seen occasional crashes on both, it's so snappy and makes such effective use of a small LCD that we're tolerating them and still using it. Given that Alpine Linux is also a useful OS for reviving sluggish old hardware, this has great potential as a happy combination. Version 3.24 came almost exactly six months after Alpine 3.23, which used the newly appointed LTS kernel, Linux 6.18. No newer kernel has since been designated as an LTS release, so Alpine 3.24 defaults to the same kernel series as its predecessor – unless you opt into following the Edge version. The Alpine Linux installation process remains quite complex and would be intimidating to newbies – although much the same can be said of Arch Linux, and that has not hindered its success, as reflected in the regular Steam surveys (although significant security issues might do). Alpine's installer has learned some new tricks in this version, including better handling of IPv6 and support for the new Limine bootloader. We upgraded our bare-metal install of 3.23, which only took a couple of minutes and went perfectly smoothly. We then installed COSMIC, which, aside from the wallpaper, looks identical to how it looks in Pop!_OS 24.04 and works just as fast. We also dug out one of our oldest ThinkPads, an X200S with a Core 2 Duo, and did a clean install of the new release on a blank SSD. It makes this 2008 laptop run so well that we'd put it into daily use if it wasn't for its flaky screen, which we forgive. We bought it secondhand a full 13 years ago, and it's seen a lot of hard work since then. The only snag we encountered with this first release of COSMIC on Alpine is that installing it doesn't add a display manager, so there's no graphical login screen, and launching it from the shell didn't work for us. When installed alongside KDE Plasma or Xfce, it works perfectly with their login screens. On a clean install, a fresh login using COSMIC took just 850 MB of RAM. For comparison, KDE Plasma 6.6.5 used 885 MB when freshly booted. Even with both desktops installed, Alpine 3.24 used a mere 3.5 GB of disk. Alpine is probably most widely used to run Docker containers, but it's more versatile than most distributions. The Reg FOSS desk mostly uses System disk mode, which is the only way most Linux distros can be installed. In this mode, Alpine runs from disk and can be set up as a normal desktop OS. It can also run in Diskless mode, in which the entire OS runs from a RAM disk, much like Tiny Core Linux. It also offers Data disk mode, in which the OS lives in RAM, but the /var directory tree is kept on disk – meaning that the RAM-based OS stores all its applications' config on disk. All of these can be customized as you wish. For instance, we have come across NAS setups in both System mode and the more radical Frood, which keeps all the software in Alpine's initramfs. If you don't mind learning some new stuff, we would definitely recommend Alpine over Arch – using musl libc means Alpine may be less compatible with some apps, but it's much smaller and faster. Flatpak goes quite some way toward helping with that, although, as we reported a week ago, Flatpak 2 might cause problems there. For now, though, if you want to avoid systemd (not to mention built-in Automatic Idiocy), then our recommendation is MX Linux for an easy life, especially if you want to dual-boot – and Alpine if you don't mind a bit more work, but prize a clean, simple, minimalist system. Xfce scales the Wayland heights As we have usually covered Xfce on Alpine – which for this vulture is a great match – we thought that for a change we'd include a few pictures of other, perhaps shinier, Wayland-based desktops. Thus, the included screenshots of KDE Plasma 6.6.5 and COSMIC 1.0.15, which are both entirely Wayland-based. The next release of Xfce, version 4.22, is expected at the end of this year. When the current release 4.20 shipped at the end of 2024, we reported that the team had begun adding Wayland support to the Xfce window manager xfwm4. At that time, it was so functionally limited that the recommendation was to use a different compositor, such as Labwc or Wayfire. That's how openSUSE Leap 16 runs Xfce on Wayland – by combining it with Labwc. This week, the Xfce project announced the first preview release of Xfwl4. It's still in its early stages and has half a dozen known serious issues, but it could be that by the time Alpine 3.25 is out, it will be possible to install it with Xfce on Wayland – without needing Alpine developer Ariadne Conill's Wayback server. Over in the Fediverse, Conill expressed her sadness that the project has not banned vibe-coded contributions – with which we sympathize. She's even mooted the possibility of starting a slop-free but Alpine-compatible distro. ®

BORK!BORK!BORK! A blast from the past greets customers at a coastal McDonald's, or is it just that the kiosk is seeking a return to the happier and simpler times of 2009? An eagle-eyed Register reader spotted a very unhappy terminal at a branch of the McDonald's fast-food chain in Worthing, England. Where a customer might normally smear a finger over suggestions for ways to sate their desire for grease, the display instead reveals the kiosk's clearly PC origins: a BIOS utility. It's not clear what has befallen the kiosk, though something has happened to the hardware that has sent it to the utility screen, which is normally accessible by holding down a key or combination of keys during the boot process. Although the BIOS is dated 2016, the system date is currently set to 2009. Worthing is a town on the south coast of England, occasionally unfairly and unkindly referred to as "God's waiting room" due to the large retiree population it once had. While the town has long since shed that sobriquet (although it still lacks the hip and trendy traits of neighboring Brighton), it seems that there is one place that would very much like to turn back the clock. McDonald's. Cast your mind back to 2009. Bitcoin was launched, an Airbus A320 ditched onto the Hudson River with no fatalities, and US President Barack Obama was sworn in. Best not to think about how much a Bitcoin acquired then might be worth now. 2009 was also the year Microsoft released Windows 7, the successor to Windows Vista and the precursor to the widely derided Windows 8. Microsoft might not have realized it at the time, but this was arguably peak Windows. Sure, Windows 95 was arguably more of a cultural "moment," and XP was an undeniable milestone, but 7 reached heights Microsoft has not matched since. 2009 was also a few years before McDonald's began rolling out touchscreen kiosks to replace the experience of peering over the shoulder of the person behind the counter to see which foodstuffs were ready to go. Today, the person behind the counter is far less visible, obscured by a line of delivery riders collecting app orders. The BIOS utility screen, however, looks back to an earlier time. Or, possibly, Worthing is simply 17 years behind the rest of the world. ®

When your accelerators sit idle, the problem usually isn't the chips. It's everything between them and the data. Rather than thinking purely about GPU performance, it's time to think about storage as an active engine for throughput, rather than a passive archive. Legacy storage architectures aren't built that way. What is GPU starvation? A starved GPU is an accelerator waiting around with nothing to do because data isn't arriving quickly enough. Sometimes the network is choking; in other cases, the next batch of training or inference data can't get off storage fast enough. Modern AI training and inference workloads demand sustained high-bandwidth, low-latency feeds that traditional storage was never designed to deliver. How do companies solve the AI storage problem? In many cases, badly. To compensate for slow, passive storage, teams copy and stage datasets into whichever environment can run the next experiment, paying what HPE calls a "staging tax" of extra hops and latency. When GPU utilization drops, those expensive accelerators become idle capital. Why does this matter now? The economics have caught up with the problem. Gartner found that only 28 percent of AI infrastructure projects fully deliver ROI. Storage increasingly shows up as the bottleneck that drives those numbers down. Pilots that ran fine on small, curated datasets hit throughput constraints the moment they scale to distributed jobs, longer training runs, and frequent checkpointing. That's where a lot of programs stall. Instead of relying on passive legacy storage, HPE advocates an "AI-ready data architecture” that gives storage the attention it needs. What does an AI-ready data architecture actually look like? Unify access first. Before chasing raw drive speed, fix the fragmentation. A unified access layer gives teams a consistent view of data across hybrid environments, so pipelines stop depending on constant copying and rehydration. Enrich on the way in. Unstructured data should arrive ready for consumption. Extracting vectors and metadata in the ingest path makes large datasets searchable immediately and exposing that metadata through open standards like the Model Context Protocol (MCP) lets agents and AI workloads discover governed data without manual tagging. Engineer for sustained throughput. All-NVMe, disaggregated designs paired with GPUDirect paths deliver data straight to accelerators and bypass the I/O bottlenecks that throttle utilization. End to end governance. Apply consistent policies, lineage tracking, and access controls across distributed data to ensure data is trusted, auditable, and used responsibly wherever it resides. What's the payoff for the business? Three things change: Iteration speeds up because engineers stop wrangling and start training. Capex stops decaying because the accelerators bought at premium rates actually run at the utilization that justified the invoice. Pilots can scale into durable production systems instead of expensive lessons. That assumes you've structured everything else in the stack correctly, from networking to model choice. The path to AI that works at scale runs through data pipelines feeding the silicon, not only through the silicon itself. Sponsored by HPE.

Train services are resuming across Germany this morning, after rail operator Deutsche Bahn (DB) last night shut down operations after its wireless network failed. At 10:30 PM local time on Tuesday night, DB advised that its GSM-R network was down, meaning all trains had to be held at stations. We understand that even suburban trains ground to a halt. GSM-R is a version of the 2G GSM standard tuned to the needs of rail operators, who use it to power private networks that carry information necessary to keep their services rolling. The tech is considered obsolete and DB knows it because the company has already signed with Nokia for a 5G replacement that will use the Future Railway Mobile Communication System (FRMCS) – a move also under consideration in the UK. For now, however, DB needs its GSM-R to connect drivers with signalling services – so three minutes after midnight on Wednesday morning the carrier promised to issue taxi and hotel vouchers to passengers. At that time, DB also said it had found the cause of the outage and was working to fix it. The company’s techies moved quickly as the network came back online at 00:50. As of 6:30 AM, however, DB warned “some isolated disruptions may still occur” and advised passengers they’ll need to check that their connections will run on time. There’s no indication the incident was the result of a cyberattack and The Register can find no reference to cut cables or other physical layer incidents that could have caused a nationwide outage. Whatever went wrong isn’t a good look as any network powering critical infrastructure is supposed to have layers of redundancy to ensure resilience. At least the org made heroes of its tech team. "Our IT experts worked tirelessly to resolve the issue – successfully,” reads a company statement. ®

Japanese telco KDDI has messed up by allowing an attacker to access systems powering an email service it manages for itself and other local ISPs, and which stores info on up to 14.2 million users. The company yesterday posted a confession [PDF] that it detected unauthorized access to the email system it offers to third-party customers on June 17th. Machine translation of the confession suggests that KDDI investigated the situation and found attackers exploited a vulnerability in third-party software used on the email service, without claiming that vuln was a zero-day it had no chance of defending or an explanation of why it was running vulnerable software. There’s some good news because KDDI was able to prevent further intrusion on the same day it noticed the attack, and says it has bolstered its defences to prevent future intrusions. But the carrier also fears that up to 14.2 million email addresses and passwords may have leaked and therefore warned that third parties may have obtained personal data. Thankfully, the company had hashed and encrypted the passwords – so users only have to fear phishing and identity theft, instead of something nastier. However, some of the data KDDI thinks may have leaked pertains to dormant accounts or others that users cancelled, meaning some potential victims will be hard to contact if the attackers have indeed stolen data. KDDI is one user of the hacked platform, and also provides it to Japanese ISPs STNet, JCOM, Chubu Telecommunications Co., Nifty Corporation, and BIGLOBE. Those companies now get to explain KDDI’s failure to their own customers, and perhaps also have the chance to revisit any other outsourcing deals with the carrier. Others who rely on KDDI to provide them with various services also get to ask the company some stern questions about whether its other platforms are secure. The carrier, meanwhile, says it’s informed the relevant authorities of the situation, but is yet to complete an investigation so remains unaware of the full extent of the mess. ®

The TOP500 list of Earth’s mightiest supercomputers has a new leader: the 2.198 Exaflop/s LineShine machine housed at the National Supercomputer Center (NSC) in Shenzhen, which took the top spot without using any kit from Nvidia, Intel, or AMD. Which is not to say that LineShine is an entirely Chinese creation. As explained in a pre-press paper, the machine’s LX2 processors are a local effort but use Armv9 designs – so chalk up a win for Blighty, the home of Arm. The machine also runs KylinOS – a Linux distribution that features contributions from around the world. The paper reveals that LineShine comprises 20,480 computing nodes, and that each LX2 processor “integrates two compute dies (304 cores total) and eight on-package HBM stacks (32 GB, 4 TB/s aggregate bandwidth).” “Each compute die contains 152 cores and 128 GB of off-package DDR memory organized into four NUMA domains,” the paper adds. “A dedicated SDMA engine handles data movement between DDR and HBM. The LX2 supports FP64/FP32/FP16/INT8 via SME and SVE units, delivering up to 60.3/120.6 TFLOPS in FP64/FP32. Nodes are interconnected via the LingQi high-speed network with a dual-plane multi-rail fat-tree topology, offering 1.6 Tb/s bandwidth per node.” That network is also a Chinese creation, from the minds at Hangzhou LingQi Technology Co. LineShine became the first system on the TOP500 to exceed two exaflops of sustained double-precision performance using CPUs only and the curators of the list think it could do better in future tests, because this time around it reached about 80 percent of its 2.736 Exaflop/s theoretical peak in tests conducted in preparation for this iteration of the TOP500 list. News that LineShine topped the supercomputing charts comes as China’s government increasingly steers local organizations towards buying made-in-China tech. Beijing wants to decrease dependency on foreign products, because China has gone all-in on AI and other technologies to boost economic growth and enhance the capabilities of its military. China’s Communist Party understands that reliance on imports can stymie those ambitions, with the USA’s ban on GPU sales to the Middle Kingdom offering ample evidence of the need to control tech supply chains. And now Beijing can point to its policies producing the most powerful single computer on the planet. It’s conceivable that China could do even better in the future, as its GPU industry is nascent and currently producing products whose performance trails Nvidia and AMD by four or five years. Those two paragons of US computing power, along with Intel, dominate this version of the TOP500 list – as has been the case for years. China is therefore on the march, but is a long way from global dominance. Our sibling site The Next Platform has extensive analysis of the TOP500 list here. ®

Modern SSDs have a limited number of write cycles before they expire. Now, OpenAI is scrambling to fix a flawed logging implementation that has been shortening the lives of Codex users' solid state drives (SSDs) with excessive data writes and lowering the devices' value by a significant amount of money. A bug report opened last week for the company's Codex coding agent warns of the consequences in its title: "Codex SQLite feedback logs can write ~640 TB/year and rapidly consume SSD endurance #28224." "On my machine, after about 21 days of uptime, the main SSD has written about 37 TB," wrote developer Rui Fan, a project management committee member of Apache Flink. "Process/file-level checks show Codex SQLite logs are the main continuous writer. "That extrapolates to roughly 640 TB/year. On a 1 TB SSD, that is about 640 full-drive writes per year. Some consumer SSDs are rated around 600 TBW, so this could consume roughly a full drive's warranted write endurance in less than a year." SSDs have a limited lifespan, often measured in terabytes written (TBW). This number varies by model and capacity. Samsung's 2025 9100 PRO SSDs, for example, promise 600 TBW for the 1 TB SSD. And after that point, we expect their performance to degrade and failure becomes more likely. The problem with Codex is that it has been writing so much logging data to SSD storage that users have become concerned they're shortening the life of their hardware. Another developer posting in Rui Fan's thread remarked, "Codex analyzed the disk usage and says this bug cost me $38.64 in drive value of my Samsung 990 2 TB NVMe." This dev subsequently cited the Codex-generated estimate of the overall cost of this bug: "This regression plausibly burned low-single-digit millions of dollars of SSD endurance across users during the March-June Window." Codex's economic impact assessment assumes a cost of $0.13 per TB written to SSDs. This is based on this formula: TB written × (SSD price / SSD TBW). So given a 1 TB SSD, we estimate that Rui Fan incurred a cost of $12.33 for 37 TB of squandered storage. (Cost per TBW = SSD price / SSD endurance = $200 / 600 TBW = $0.333 per TB written.) A more spacious and more costly SSD with a higher TBW rating would cost less per wasted byte (e.g. $0.25 per TB for a $300 / 1200 TBW 2 TB Samsung 9100 PRO SSD). In December 2025, Codex devs announced plans to add telemetry by default (except where disallowed by law) to the Codex CLI. But this issue has to do with local diagnostic logging, which was introduced around the time the app debuted last year and is also on by default. The logs stay on the device unless included by the user in a feedback report. Concerns about excessive write operations using OpenAI's Codex have been surfacing in the project's GitHub repo for several months. A spokesperson for OpenAI confirmed that company engineers are aware of the problem and are working to fix it – something evident from several recent pull requests intended to address the problem. We're told that these logs are intended to help OpenAI engineers diagnose issues and that the problem was the result of high-volume data that was being stored in a way that created far more disk activity than anticipated. While purported fixes have been landing and the company has made some progress, users continue to file problems. The issue appears to date back to work done in February to write app-server SQLite logs at TRACE level, which emits more verbose logs than, say, ERROR level. We note that Codex, presumably running GPT-5.3, reviewed this particular series of commits. That makes it all the more surprising that the code was so ill-conceived. ®

Anthropic is killing off its existing Claude in Slack app in favor of an always-on agentic alternative dubbed Claude Tag that will listen and learn from everything it’s given access to. Claude Tag can now join organizational Slack instances “as a team member,” according to Anthropic’s announcement, where it will have access to whatever channels domain administrators decide upon, as well as the tools, data, and codebases they contain. Anyone in a channel with Claude Tag can tag @Claude to delegate tasks to the bot, which will be able to perform a variety of actions and build context about its environment based on what flows through a channel. “We see Claude Tag as the beginning of an evolution of Claude Code,” Anthropic said. “It makes the model even more proactive, and it works better with a full team.” Anthropic further explained that it’s using Tag to write code (64 percent of what comes from the product team is apparently written by Tag) as well as “chase down product metrics and data, work through support tickets … [and] help find the root cause of tricky bugs.” Integrating Claude into Slack isn’t exactly new – there’s already an existing Slack connector app for Claude, but it’s being depreciated in favor of Tag. According to the Claude in Slack help page, the existing connector will leave service on August 3, and Enterprise and Teams customers are getting access to Tag beginning today, per Anthropic’s announcement. The persistence and contextual learning that Tag brings are its major changes over the prior Claude-Slack connector, according to Anthropic. Tag is shared across a channel, meaning that everyone who interacts with it in a particular space will interact with the same Claude as their coworkers, “more like interacting collaboratively with a teammate” instead of just getting one-off advice from a bot, according to Anthropic. As mentioned above, Claude Tag will also learn the longer it’s in a channel, meaning users won’t have to explain things to it every time they chat with it. A Claude Tag based in one channel can also learn from other channels and data sources, provided that it’s given permission to do so. Beyond that, Claude Tag can also take initiative to “keep you updated about whatever it thinks you might need to know” across channels and tools it has access to, as long as ambient behavior is toggled on. Anthropic says that it’s also able to schedule tasks for itself in order to autonomously pursue its own assignments “over hours or days.” Anthropic’s announcement said that its employees have found that feature particularly helpful, as instead of doing their own work “we now spend much more of our time delegating tasks to many Claudes in parallel,” which is supposed to sound like a good thing. Act now; free credits running out and you'll need 'em Administrators of Slack environments using the old Claude for Slack app have 30 days to opt in to the Tag migration. That said, Anthropic is still keen on keeping those user numbers up, so it’s offering a hefty volume of credits to anyone on an Enterprise or Teams plan with more than 10 total seats as long as they migrate before the credits expire on September 1. According to a support page, Enterprise customers who add Tag to Slack will get $25,000 worth of credits, while Teams customers will get $2,500. Those credits are only applicable to Tag usage in Slack, but not for direct messages with Claude, which are billed to individual seats. AI token use has recently become a hot topic, especially since Anthropic did away with bundled usage pricing in favor of metered pricing that has surprised many customers with large bills. It's not clear how far those Tab credits will stretch, and Anthropic didn't respond to questions for this story. It's also worth nothing that using Tag means handing business data over to an AI model that lives on Anthropic’s servers, so caveat emptor to anyone adding this agent to channels where sensitive info may be discussed and ingested for contextual learning. Then again, Slack is shunting potentially sensitive business data into the cloud too, and it’s not always safe with the Salesforce subsidiary, either. ®

Sometimes it takes a while to detect a vuln. A 29-year-old, Heartbleed-style vulnerability in Squid, a popular open-source caching proxy server, silently leaked users' plaintext HTTP requests and potentially revealed sensitive data, including credentials and session tokens, for decades - until AI (and a few humans) saved the day. A security researcher and Mythos Preview found the flaw and reported it to project maintainers, who fixed the code earlier this month. Squid is widely used by large corporations, schools, and internet service providers to cache, filter, and monitor network traffic, and Calif.io researcher Lam Jun Rong said he came across the open source proxy while attempting to connect to the internet on a flight. “As you might expect, the version of Squid deployed on that plane was released nearly 10 years ago and is affected by the vulnerability I'm about to share with you,” Rong wrote in a blog post about the bug, which he dubbed Squidbleed and investigated with help from Anthropic's Claude Mythos Preview. Rong reported the bug, tracked as CVE-2026-47729, to Squid’s maintainers back in April, and it’s fixed in Squid v7.6, released June 8. The Reg readers may remember Calif from their earlier HTTP/2 Bomb research, uncovered by OpenAI’s Codex agent, and the AI bug-finding firm also collaborated with OpenAI on its Patch the Planet initiative, announced on Monday. According to Rong, Squidbleed leaks internal memory from every version of Squid in its default configuration with two conditions. First, Squid has to be able to read and inspect the network traffic, so it must be handling cleartext HTTP (not HTTPS) or be deployed in TLS-terminating setups. Additionally, the proxy must be allowed to reach an attacker-controlled FTP (File Transfer Protocol) server via TCP port 21. FTP is an outdated protocol for moving files between machines, and Squid supports it - which is where the problem lies. The bug exists in Squid's FTP directory listing parser, and it was injected into the open source code as a commit (bb97dd37a) created in 1997 to support old NetWare servers. NetWare is a discontinued network operating system that was popular in the 1980s and 1990s, providing file and print services across local area networks before Windows and Linux servers became dominant. NetWare FTP servers also added extra whitespace between the modification timestamp and the filename, compared to most other FTP servers that just used a single spFace. The 1997 commit fixed this NetWare issue by instructing the code to skip the extra whitespace using this loop: while (strchr(w_space, *copyFrom)) ++copyFrom;. As Mythos Preview discovered, if an attacker's FTP server doesn't provide a filename after the modification timestamp, copyFrom points to the terminating NUL character at the end of the string. “strchr treats that terminating NUL as part of the string it searches, so it returns a pointer instead of NULL, and the loop never stops,” Rong explains. “It walks off the end of the buffer, and xstrdup copies whatever follows back to the attacker as a filename.” This results in a heap overread and can leak HTTP requests that often contain passwords or API keys, and Rong demonstrated this exploit in a proof of concept. “The patch is simple: check for the null terminator before calling strchr,” Rong wrote. If you use Squid, make sure to download the June release to fix this flaw. Also, as Rong suggests, you should disable FTP unless there’s a “specific, unusual need for it.” Chromium-based browsers stopped supporting FTP years ago and for good reason. This means “most organizations running Squid are getting close to zero legitimate FTP traffic,” the security sleuth noted. “Turning it off removes this entire attack surface for free.”®

Rocket Lab has just completed one of the most rapid space launches ever, kicking off a complicated exercise that will test the US Space Force’s ability to respond to and characterize potential threats in orbit. The Victus Haze mission, as the project is known, saw Rocket Lab lob one of its Pioneer spacecraft into orbit for the Space Force on June 19 in just 16 hours and 42 minutes, beating the notice-to-launch record set during the 2023 Victus Nox tactically responsive space mission (TacRS) by more than 10 hours. In addition to the rapid launch, Rocket Lab also managed to fully ready its Pioneer spacecraft being used for the exercise in just 37 hours and 36 minutes, well in advance of Victus Haze’s 72-hour commissioning deadline. With Pioneer now in orbit, the next phase of Victus Haze is set to begin. “The mission will now transition into an on-orbit focus placing operationally relevant systems through realistic rendezvous and proximity operations (RPO) threat response scenarios,” the US Space Force said in its own announcement of the Rocket Lab launch. Pioneer won’t be flying RPO maneuvers on its own, though. Rocket Lab may be playing the part of the Space Force’s rapid launch partner for Victus Haze, but space defense company True Anomaly already has one of its Jackal satellites in orbit. Jackal, which was launched on a SpaceX Falcon 9 rocket in May, will be playing the role of a “non-compliant satellite” that Pioneer will have to rendezvous with and characterize, demonstrating the ability for both craft to maneuver around each other, take photographs for analysts to pick apart, and track each other as if they were hostile targets the USSF wanted to monitor. As noted by True Anomaly, Victus Haze marks a departure from Victus Nox, which only involved a single spacecraft and focused on uncontested orbital operations. “VICTUS HAZE encompasses the full scope of TacRS operations: rapid launch and initialization, followed by operationally relevant 1-on-1 RPO between Rocket Lab and True Anomaly spacecraft in low Earth orbit,” True Anomaly said in its own announcement of the Victus Haze mission. Those one-on-one maneuvers the two satellites will undertake will help develop tactics, techniques, and procedures for future space operations, True Anomaly explained, as well as determining what sort of equipment the Space Force might want to consider for its TacRS spacecraft. The mission marks the second of the USSF’s planned annual tactical space missions, a pace that has already slipped after no TacRS launch took place in 2025. Victus Nox, the first full-fledged TacRS mission involving an actual space launch, launched in 2023 and concluded in 2024. Victus Haze’s notice-to-launch record comes a week after DARPA announced it was seeking concepts for rapid-launch space missions able to quickly replace space assets destroyed in an orbital conflict, again suggesting that star wars - or at least orbital ones - are rapidly looking inevitable. ®