Despite previous controversies, it seems like the lords of Linux are warming to AI-assisted coding. The penguin's kernel community has agreed to stop asking coders to disclose whether they used bots to help with their contributions. And Linux kernel supremo Linus Torvalds has made clear that he opposes "stupid" AI disclosure rules as long as people understand the code they're submitting. "The Maintainers Summit reached consensus to deprecate the Assisted-by Tag," said Sasha Levin, a Linux kernel developer and engineer at Nvidia, in a message to the Linux kernel mailing list last week. "Remove the documentation requiring it." The policy had said that those using "any sort of advanced coding tool in the creation of [their] patch" should acknowledge that through the addition of an "Assisted-by" tag in the commit message. The AI labeling proposal initially took shape in July 2025. It became part of the documentation in January 2026. This was shortly after Torvalds said that he was fairly positive about vibe coding as a way for people to get into programming, but remained skeptical about its effect on software maintainability. Since then, he's warmed to AI tool usage. AI attribution wasn't an issue when the generated code was obvious, buggy, and bad. But in the past year or so, AI coding agents started producing code that is of acceptable quality for the purposes of the Linux kernel community. "LLMs are good enough now," said Greg Kroah-Hartman, a veteran Linux kernel maintainer, at the Linux Foundation's Open Source Summit in South Korea in August. "They've gotten good enough to find bugs because LLMs, AI, are just pattern matchers and code is nothing but patterns in a specific format." Cautious acceptance hasn't erased concerns about errors slipping through unnoticed, accountability, and coding agents reproducing copyrighted code. There's also still a stigma in some circles around the use of AI tools. Those who have spent years honing software skills, coding with limited assistance from tools, may disparage developers who don't understand the code they conjure with a prompt. Rapid change will do that – in the space of a few years, software production practices have changed drastically, and software quality, security, and maintenance practices have yet to fully adjust. The leading makers of AI coding agents – Anthropic, Google, Microsoft, OpenAI, and the like – are doing their best to promote AI-driven projects as better than anything a human developer could manage. Yet even they differ in their AI attribution practices. Claude adds provenance messages by default, but Codex does not. In fields such as advertising and art, content provenance can help assess misinformation and human creative input. For code, the focus tends to be more on functionality than form. In a prior discussion of the merits of the tag requirement, Jeff Layton, a Linux kernel developer with Red Hat, explained the rationale for reversing course. First, he said, not everyone abides by the policy. Second, he said, "it basically serves as free advertising for proprietary LLM companies." And third, he said, "it's not clear why we want to collect this info in the first place." Some kernel contributors still see value in the tagging requirement. "I'm not happy removing the tag without adding another mechanism for required disclosure," said Laurent Pinchart, a Linux kernel dev and founder of Ideas on Board, in a mailing list post on Tuesday. "Without that, I can foresee submitters pushing back on the question." Pinchart challenged the argument that AI labeling rules will just be ignored. "What this is about is setting our expectations, telling what we want and what we don't want," said Pinchart. "There's still value in that, and I'd argue there's even more value now. There will be people who won't care, there will be people who cheat. We've always had those, we'll keep having them, and we'll keep rejecting their contributions. I really don't get the unenforceable argument here." Linus Torvalds responded with an emphatic rejection of tool usage declarations. He wrote: What you are about is something entirely different. It has nothing to do with "expectations", and it has everything to do with your hangups. I don't understand this whole idiotic concept of "don't use AI for X". Here's a new rule just for you: don't use compilers for any of your kernel submissions. Because if you can't show that you really understand what the machine is actually doing by writing your code in assembler, then I don't want to see your code contribution. You clearly aren't competent if you can't write it that way. And btw, you had better not use an editor to write that commit message either. Because I need you to show me that you can actually write a good commit message without using tools that make it easy to backspace and hide the mistakes you made while typing. Ok? Because otherwise those tool-assisted commits of yours clearly aren't showing that you know what you are doing. Do you not see how UTTERLY STUPID those kinds of rules would be? So no. We're not going to make stupid rules that say "you can't use tools for X". The kind of expectations *I* have is that people do their best and I expect people to understand what they are doing. But that understanding has very little to do with what tools they use. And quite honestly, in my experience, the people I respect and call intelligent are not the ones that shy away from using new and interesting tools. ®

Researchers at security startup Xbow used OpenAI's GPT-5.6 Sol model to help chain three zero-day vulnerabilities in Chrome into an exploit capable of executing code outside the browser's sandbox on Windows machines. Xbow estimates that this could affect about one billion users worldwide. Google fixed all three, and paid the Xbow team a $250,000 bounty, its top reward for a Chrome full-chain exploit. The payout also proves a nice return on investment for the human bug hunters - Brendan Dolan-Gavitt, Andres Luksenberg, Baul Lee, and Willem Pinckaers - who spent just six days during the Black Hat and DEF CON conferences and about $40,000 in token charges to discover the vulnerabilities and build the attack chain. “The number on the check is impressive, but the time it took our team, with the help of a publicly available AI model, to break in is even more impressive,” Xbow CEO and founder Oege de Moor said on Tuesday. “A three letter intelligence-agency veteran told me that his team used to spend six months trying to identify this kind of vulnerability and get total access. We did it in six days.” In a separate technical writeup about the attack chain, Dolan-Gavitt detailed the work done by humans and AI agents. Agents, using OpenAI's GPT-5.6-Sol model, searched source code, tested the team’s hypotheses, and developed proofs of concept and exploit stages, he said. Meanwhile, the AI pen-testing security startup’s researchers chose targets, provided context, challenged the agents’ results, and decided which attack components to chain. "We did apply a fair bit of our own human expertise and judgment," he wrote. Crossing two security boundaries in 3 steps Google’s Chrome browser uses several layers of security to protect users, and this includes sandboxes - barriers to contain compromised code and ensure that someone who hacks an individual website can’t take over the entire computer. This includes deploying the V8 sandbox to prevent any memory corruption flaws in the V8 JavaScript engine from impacting other memory. The V8 sandbox sits inside the operating-system process sandbox, and this nested design means that escaping V8's memory restrictions doesn't automatically grant native code execution. Even if an attacker achieves native code execution inside Chrome's renderer, the outer process sandbox restricts access to the rest of the machine. “The chain crossed both boundaries in three vulnerability stages,” abusing Chrome to run code on Windows machines, Dolan-Gavitt wrote. The attack combines a V8 compiler bug, an animated-WebP decoder bug, and a memory-safety bug in Crashpad. The first vulnerability is in Maglev, V8's mid-tier optimizing compiler, and it’s due to a reused, mutable boxed number during deoptimization. This gives attackers read and write access throughout V8's pointer-compression cage. The second stage, the animated-WebP bug, is a V8 sandbox escape. “But rather than attacking the sandbox itself we exploited a property that the V8 sandbox deliberately does not enforce and that Blink incorrectly assumed it could trust,” Dolan-Gavitt explained. “Finally, the exploit used an overflow in Crashpad to cross the final outer process boundary and achieve unsandboxed code execution on the host.” While Google patched all three security issues, and Chrome installs updates on its own, a fix only takes effect once the user relaunches the browser. If you use Chrome, make sure you are running version 152.0.7977.82 or later. The time and money spent to achieve a full chain Chrome exploit illustrate - yet again - that “the bar is significantly lower” for attackers armed with LLMs, Dolan-Gavitt wrote. “Defend accordingly.”®

Ivanti customers are being urged to update Connect Secure and check for compromise after attackers exploited a newly disclosed zero-day vulnerability. On Tuesday, Ivanti released updates for Connect Secure (ICS) to fix CVE-2026-104935, an authentication bypass flaw discovered by Volexity. Rated 8.6 on the CVSS scale, it allows remote, unauthenticated attackers to execute arbitrary commands. "We are aware of exploitation in the wild affecting a limited number of customers at the time of disclosure," the vendor said in its advisory. There are limited indicators of compromise available at present, it said, advising customers to refer to Volexity's report on the findings for more information, and to run the Ivanti Integrity Checker Tool regularly. Ivanti lists affected releases in the following branches: 25.1.1.1 and earlier 25.1.2.2 and earlier 22.8R2.4 and earlier The vendor said it previously identified "a portion of this vulnerability" and patched it in July's 25.1.3.0 update, and had been working on backporting the same fix since then, however, attackers began exploiting it before that work was complete. Volexity said the attackers behind the recent intrusions were previously unknown to its researchers, but are now tracked as UTA0576. Criminal motive rather than espionage is suspected by Volexity. Based on observed post-exploitation activity, attackers appeared to be most interested in using their newfound access for credential theft, data exfiltration, log tampering, and network reconnaissance. Notably, the attackers did not deploy any malware nor execute web shells, using only the commands available to them on the ICS appliance. "Because the operation was detected and disrupted, the threat actor's ultimate objectives could not be determined," said Volexity. "However, Volexity assesses with high confidence that UTA0576's actions signaled preparation for lateral movement and a broader compromise of the victim network." The researchers also suspect that the attackers were using large language models to support their endeavors. Volexity pointed to the attackers' repeated use of the exploit channel to run commands, rather than deploying web shells, and their use of "highly specific scripts using hardcoded credentials." Ivanti disclosed one CVE, but Volexity described an attack chain involving three vulnerabilities. Tuesday's updates address the authentication bypass and remote code execution stages. The third vulnerability, which allows privilege escalation, is separate from the flaws encompassed by CVE-2026-104935. Volexity said Ivanti is aware of the additional bug and plans to publish a fix "in the coming weeks." The disclosure of the zero-day this week makes it the third year running that Ivanti has had to deal with attackers exploiting undisclosed flaws in Connect Secure. The previous ICS zero-days were disclosed in January 2025, a year after the January 2024 vulnerabilities led to exploits affecting thousands of ICS appliances. In the interim, we've also seen suspected Chinese espionage groups exploiting bugs in Endpoint Manager Mobile (EPMM), and critical flaws in Sentry. Ivanti's run of January zero-days continued earlier this year, although those flaws affected EPMM rather than Connect Secure. ®

It's about time: the General Conference on Weights and Measures looks set to end leap seconds in an effort to reach horological harmony. Since 1972, 27 leap seconds have been added to Coordinated Universal Time (UTC) to keep it in step with the astronomical time scale UT1, which is based on Earth's fluctuating rotation. The proposal would increase the permitted difference between UTC and UT1 from 0.9 seconds to one hour, avoiding leap-second adjustments for centuries. Earth's rotation has also sped up in recent years, raising the possibility of a negative leap second – removing a second rather than adding one. That has never been done before. Experts have estimated a 30 percent probability that global systems will need a negative leap second by 2035 to keep the two time measures in step. That could pose problems for computer systems designed to handle added seconds but not removed ones. Many national bodies support the proposal, which is expected to pass. If approved, the change would take effect on May 20, 2027. Setnam Shemar, principal scientist at the UK's National Physical Laboratory, said the proposal would avoid the risks associated with leap seconds for centuries. "Leap seconds can be difficult to implement reliably, and if not applied correctly, they can cause differences in time across networks, potentially leading to the failure of systems such as mobile phone networks or electricity grids. These risks could increase if a negative leap second had to be implemented for the first time. "Previously, leap seconds were added to UTC to keep it within 0.9 seconds of UT1. From 20 May 2027, a difference of up to 1 hour will be allowed before intervention is needed, ensuring the continuity of UTC for several centuries or more. "The relationship between UTC and UT1 is important for some applications, such as astronomy and celestial navigation. The decision would not sever that link: UT1 will continue to be monitored, and its difference from UTC will continue to be disseminated through established international services." ®

Silicon Valley’s AI bros have supposedly found a new way to signal to those around them how committed they are to their industry: Take up smoking despite the risk of cancer, because they expect bots to develop a cure … any day now. At least, that’s the claim Palantir CTO Shyam Sankar made in a recent interview on CBS's Face the Nation. In trying to explain what he described as a doomsday cult surrounding effective altruism, whose adherents include some AI researchers who believe the technology could bring about a utopia, Sankar claimed smoking has become a way for researchers to demonstrate their faith in the tech. “So many AI researchers have started smoking as a way of signaling to other researchers how much they believe in AI,” Sankar said to the confusion of host Margaret Brennan. Those researchers “believe so much in AI that it's going to cure cancer, I'm showing you my chips are all in,” Sankar explained. Sankar described the decision to take up smoking as a bet that AI would cure any health ailments that arose from it as “mass psychosis” and “social delusion” that’s getting in the way of actual discussions around the state of AI, and he blamed Silicon Valley’s pet philosophy, effective altruism, for those delusions. “There's an unseen force driving the narrative around AI safety, and it's really this millenarian doomsday cult meets messianic complex called effective altruism,” Sankar said. That philosophy prioritizes making the most effective decisions that benefit the most humans – like utilitarianism on steroids – and it tends to take a long-term approach to things. That’s where AI comes in: Effective altruists are ostensibly focused on ensuring AI systems are built in a way that will best benefit humanity. That focus has been a key reason why so much AI doomerism has come out of Silicon Valley too: If we’re going to build safe AI, says the effective altruist, we have to make sure that we consider all the risks. “There's only really roughly 2,000 AI researchers in America, and the vast majority of them believe in this sort of worldview around effective altruism,” Sankar noted. The Palantir CTO seems to believe, like many effective altruism critics, that all the doomsday scenarios and far-fetched risks being posited by the EA AI crowd are less technical critiques and more like theological ones. “I think there's a profound truth to Pascal's saying that every human has a god-shaped hole in their heart,” Sankar said, “and it's important to understand that many of these folks in Silicon Valley who are building this technology have filled that hole with the machine God of AGI.” Smoking as a bet that the machine god will save them from cancer, therefore, is akin to the ascetic’s endurance of physical pain or trauma in the hopes God will heal them or see them safely through their torment. From a more practical perspective, this isn’t exactly a great idea, as one former cancer researcher who spoke to The Register told us. “Good luck to them with this approach,” the researcher, who declined to be named, opined. “Even more so, good luck to their non-smoking friends and family taking in their passive smoke.” Taking up smoking in the hopes of a cure may not be a wise idea when one considers the timeline needed to develop a cancer drug, much less a cancer cure. "Even with AI, developing and introducing a new cancer drug into any of the world's health systems can take several years,” the researcher explained. “Even then we wouldn't know whether or not resistance to the drug would develop in the tumors.” The ability of AI to discover drugs and medical treatments has been much hyped – Google DeepMind CEO Demis Hassabis even predicted last year that 2025 would be the year humans started popping AI-designed pills. That hasn’t happened yet, and there’s no real indication it will anytime soon. In other words, the likelihood that AI is going to kill us all is ridiculously overblown. The actual casualties, it seems, are more likely to be the AI bros giving themselves cancer as a show of faith in their machine god. ®

A browser-based emulator for the original iPhone is the latest project to tap into the rich vein of nostalgia running through the tech industry. While far from a complete implementation, the emulation does enough to evoke happier times, before Apple burdened its mobile operating system with frills and complexities, and spent years (and multiple operating system updates) making the Podcasts app steadily more broken. The browser-based emulation (and we use the term carefully; anyone expecting all the functionality of yesteryear will be sorely disappointed) shows an original iPhone as it debuted in 2007. The buttons are where you remember them, the screen is ringed by bezels that wouldn't pass muster today, and the skeuomorphic icons are all present and correct. There's even the browser, though it doesn't do much yet. Even a gentle prod shows how incomplete this implementation is. Some apps work well, such as Calculator, while others - Contacts - struggle, and anything that needs an online connection fails, including loading today's web in 2007's Safari. It does, however, recall how simple things were almost 20 years ago, when the iPhone first appeared, and it joins an array of old-time systems that can be fired up (or at least look like they are) in a browser. Want a look at Apple's abandoned System 7 replacement? Step this way. Perhaps an XP-like implementation? No problem. It would be churlish to dwell on the emulation's deficiencies. For now, it's arguably more curiosity than anything else: Apple didn't launch the App Store until 2008, so third-party apps would be off the table, at least until a later version of iOS landed. Running it in a browser is undoubtedly convenient, though. Alternatives such as Light Touch offer a fuller implementation, but they demand more than a handy Chrome instance on a Windows device to open a portal back to a brighter time. So, fire up 2006's Nintendo Wii, pop on that CD of Back to Black, and relive the excitement of switching on the very first iPhone. Although perhaps give that other 2007 icon, Windows Vista, a bit of a swerve… ®

Gigabyte recently announced new motherboards that use DDR4 RAM, a standard that launched in 2014, and other vendors support even older standards in brand-new kit. Earlier this year, AMD released a Ryzen 7 5800X3D anniversary edition to mark ten years of AM4. For AM4 processors, Gigabyte offers the new B450M H ARGB, with two DDR4 slots, and the B550 X-Eagle with four slots. Those who prefer Intel chippery can now buy a B760M Aorus Elite. This is notable because of the age of the DDR4 spec. In case you lose track of these things, The Register first described the new DDR4 spec back in 2012, and a few months later urged the keen to wait a year or two. DDR5 supports higher transfer rates, and The Reg predicted its arrival in 2017. The DDR5 standard was agreed in 2020, with mainstream PC adoption following in 2021. Over in the People's Republic of China, several vendors will sell you new DDR3 motherboards for 15-year-old server CPUs, such as Huananzhi's X79-4D or X79 Supreme, or the Machinist X79 V2.82H. These use Intel's LGA-2011 socket – named for the number of pins, but coincidentally launched in 2011. This Chinese SMZDM community discussion is about whether the circa 15-year-old Xeon E5 is still worth buying – a CPU The Reg covered in 2012. This is partly due to the current price rises and component shortages induced by large-scale, cloud-based LLMs – but also "training" those "open-weight" models that run on your own kit. Those too are constructed in giant datacenters stuffed full of GPUs, ingesting all the human-written material they can get their hands on, rarely asking for consent. And not just stuff put online either. Once you wipe the sluggish, bloated commercial OS from your computer's storage, you discover two things. First, you don't need the latest hardware any more. Secondly, the older the PC (or Mac), the better the chances that it will work fine with Linux – or indeed the other FOSS OSes out there. Current and recent hardware is the most problematic for FOSS developers because most manufacturers prioritize Windows testing, while restricted documentation and proprietary drivers can complicate open source support for new hardware. Together, these things mean that if you're happy without high-powered local LLMs, or three-digit frame rates, second-hand kit is more usable than ever. (For those who can't do without one or two Windows-only apps that won't run under Wine, we remind you that desktop VMware is free now… and for better desktop integration, we looked at WinApps and WinBoat earlier this year.) The Reg FOSS desk is based in the British Isles, close enough to order stuff inexpensively from the UK. We've been using online vendors of depreciated ex-corporate kit for decades: we've bought several machines from Morgan Computers, and occasionally from Tier1 Online and Bargain Hardware. Other vendors are available. There are also special-purpose price-comparison sites, such as LowCostMiniPCs.com, which searches other sales sites for bargains. When we were based in Czechia, we bought a few things from GigaComputer, and we know of others we didn't try, such as Refurbed.cz. Other such European sites include Computershop.eu. The usability of kit that can only officially run Windows 10 was the entire point of the End of 10 campaign. Microsoft cites security benefits for Windows 11's hardware requirements, but in this vulture's opinion, those benefits don't justify excluding an entire generation of otherwise capable hardware. Keeping older hardware in service is welcome. If digital sovereignty gives organizations another reason to consider open source alternatives, so much the better. ®

Tiny11 Builder, a tool for creating less-bloated Windows 11 installation images, now has a desktop app preview that makes the process less reliant on PowerShell know-how. As the author explains: "Tiny11 Builder aims to be the de facto standard for creating lightweight, safe and efficient Windows images." The tool creates an ISO image for Windows installation, stripping out bundled apps and selected system components. It accepts Windows 11 installation media from 22H2 onward, including 26H2, and the result is something a good deal more compact than Microsoft's original. The major visible change is a WinUI 3 desktop application that uses a new C# image builder rather than merely running the existing PowerShell scripts behind a graphical interface. Those scripts remain available, but the new interface is considerably more polished. There are three basic build modes: Tiny11, which removes bundled apps while retaining Windows Update, Defender, and recovery tools; Tiny11 Core, which disables updates and Defender and irreversibly removes components, including the recovery environment; and finally a custom option where users can decide which Windows elements to keep or drop from the ISO image. Tiny11 remains an impressive technical achievement, but it carries an important health warning: the ISOs it creates are definitely not supported by Microsoft, and if you opt for Tiny11 Core, the author recommends sticking to test VMs only. This is for enthusiasts who understand that a smaller installation can come with missing functionality, security trade-offs, and support complications. There is also no guarantee that attempts to bypass some of the more onerous aspects of Windows 11 installation will work. As the author notes in the project's repository: "Local-account and hardware-check settings are best effort, not guaranteed bypasses on new builds." In other words, don't count on it to bypass Windows 11's hardware requirements or account demands on newer builds. Tiny11 is all about stripping fluff and fripperies from the Windows 11 installation image. As with earlier versions, a license is still required. While the ISO might have been debloated, this is still Microsoft Windows 11 after all. "No modification bypasses Windows licensing or activation," the author notes. The appeal of removing Microsoft's excesses is clear. Whether the reduced footprint is worth the loss of components and potential support headaches depends on what users choose to strip out. ®

AI datacenters are a rats nest of fiber optic cables stitching together thousands of AI accelerators, NICs, switches, and everything else required to train and run the latest and greatest modes. Cisco hopes to untangle this mess with a new pluggable optic developed in collaboration with Nvidia that will pack between four and eight times the bandwidth of existing pluggables into a single port. Announced at the Open Compute Project (OCP) global summit this week, the ultra-dense small form factor pluggable (USFP) promises to pack up to 32, 400 Gbps lanes into a pluggable roughly the size of existing OSFP modules. Each of these pluggables would boast 12.8 Tbps of Ethernet bandwidth, as much as a high-end enterprise switch just a few years ago. To put that in perspective, the most sophisticated pluggable optics today cap out at around 1.6 Tbps using eight 200 Gbps links. Such a pluggable has the potential to significantly simplify cabling across AI datacenters where it isn't uncommon for each GPU to be paired with an 800 Gbps and soon 1.6 Tbps NIC. A single USFP pluggable could consolidate eight GPUs worth of cabling into a single fiber optic bundle. Meanwhile, on the switch side, a future 204.8 Tbps switch would only require 16 USFP ports to support 128 1.6 Tbps endpoints. Cramming all that connectivity into a single pluggable won't be easy. Today a high-end pluggable optic might pull on the order of 20 watts. Cisco's USFP spec calls for up to 100 watts per pluggable. There are also design constraints that have to be taken into consideration. Four times as many links per pluggable means 4x as many electrical contacts that have to be perfectly seated. Because of this Cisco says USFP modules will employ a two phase insertion mechanism in which the pluggable is first inserted into the cage and then a second mechanism is actuated to make the electrical connection. Speaking of reliability, USFP would have a bigger blast radius than conventional pluggables. A single bad laser or digital signal processor (DSP) has the potential to take out far more than a single GPU, but entire systems. While that might sound bad, it's nothing compared to co-packed or near-package optics where one failure could drop an entire rack's worth of accelerators from the network. Nvidia's participation in the development of the USFP standard is notable as the Ethernet networking titan has largely moved away from pluggables in its most recent set of switches. Many of its Spectrum-X and Quantum-X switches now use co-packaged optics and for good reason. Cutting out the pluggables on the switch side nets a roughly 30 percent power savings. But while Nvidia's latest switches may use CPO, its NICs still rely on pluggables in part because it offers integrators greater flexibility. Not everyone is building 100,000-plus GPU clusters for training. For racks utilizing a top of rack switch, lower power direct attach copper cabling can be used without much issue. Given Nvidia's investments in CPO we can only speculate on its interests in USFP, but it likely comes down to densification. The AI infrastructure giant is rapidly moving toward 600 kW racks packing 144 GPUs on board. First teased in early 2025, Nvidia's Vera Rubin Ultra platform promised 115.2 TB/s scale out bandwidth per rack or about 3.2-6.4 Tbps of bandwidth per GPU package depending on whether they're counting bidirectional bandwidth or not. It just so happens that Cisco's initial USFP spec calls for 16, 400 Gbps links per pluggable, or about 6.4 Tbps of bandwidth. While Cisco and Nvidia may be leading the charge on the development of an ultra dense optical pluggable, like most things Ethernet, USFP will be an open standard anyone can use in their designs. Microsoft taps Cisco for multi-datacenter AI Alongside Cisco's latest tie up with Nvidia, Switchzilla also revealed that Microsoft is now using its Silicon One P200 routing chips to distribute workloads across multiple datacenters. Nvidia and others have taken to calling this kind of multi-site computing: scale across networking. We discussed Cisco's datacenter-scale routing platform at length late last year, but the silicon competes directly with products like Nvidia's Spectrum-XGS and Broadcom's Jericho 4-based offerings. Cisco touts the P200 as the first 51.2 Tbps systems "deployed in a production environment by leading hyperscalers." With enough qualifiers you can be the first at anything. As we reported at the time, Microsoft was among the first major hyperscalers to embrace scale-across networking as a means of overcoming local power constraints with its Fairwater clusters in Wisconsin and Georgia. And it seems we finally know the hardware good old Redmond is using to stitch them together. Cisco gives SONiC some love Speaking of hyperscalers, one thing you probably won't find in a hyperscale network is Cisco's Nexus network operating system (NOS). Hyperscalers like Microsoft typically employ vendor agnostic NOS platforms to avoid vendor lock-in and make managing networks across multiple Ethernet vendors cheaper and easier. Originally developed by Microsoft and later open sourced, SONiC is among the most popular and widely deployed of the hyperscale NOSes. SONiC support, we should note is nothing new for Cisco which contrary to popular belief is more than willing to sell you any one piece of its networking stack, assuming you've got a big enough purse — the hyperscalers certainly do. As such Cisco is one of the biggest contributors to the SONiC project. Instead of forcing customers to use its software stack, Cisco's approach to SONiC and the hyperscale is to extend the NOSes capabilities to better take advantage of ASIC's unique capabilities and allow for integration with its services. At OCP this week, Cisco announced support for its Live Protect platform in SONiC to make it easier for customers to keep up with security patches in the wild. This capacity is expected to roll out next year. ®

Some of America's biggest tech companies have a warning for Europe: curb your reliance on US cloud providers, and it could cost you €115 billion. That's the finding of a new study commissioned by CCIA Europe, a lobbying group whose members include Amazon, Google, and Meta, companies that have plenty of reasons to want European governments to keep buying American cloud services. The research, conducted by the European Centre for International Political Economy (ECIPE), claims the European Commission's proposed Cloud and AI Development Act (CADA) could saddle Europe with €115 billion in costs over the next decade, more than six times the Commission's upper estimate of €18 billion. The report argues that restrictions favoring European cloud providers would leave governments and businesses facing hefty bills to migrate workloads and rewrite software. It puts direct public-sector costs at €33.3 billion, with a further €80 billion potentially falling on private companies in sectors including banking, healthcare, energy, and transport. However, a rather large asterisk is attached to that €80 billion figure. It assumes the Commission extends sovereignty requirements beyond public procurement to cover 25 percent of cloud workloads in essential private-sector industries. In other words, it's a modeled scenario, not a bill Brussels has already handed European businesses. The study also comes with some baggage of its own. CCIA represents several US technology giants that dominate the European cloud market, including Amazon and Google. Both have spent years selling European customers cloud services while more recently developing sovereign cloud offerings designed to address concerns about foreign ownership, data access, and regulatory control. Amazon, for example, has been developing its AWS European Sovereign Cloud, while Google markets a range of sovereign cloud solutions. Apparently, sovereignty is an attractive proposition when it comes with a hyperscaler logo attached. Meta, another CCIA member, has its own interests in European AI infrastructure, although it is not a commercial cloud provider on the scale of AWS or Google Cloud. The report specifically targets Article 31 of the proposed legislation, which it says could allow the Commission to extend sovereignty requirements into essential private-sector services. CCIA argues that forcing businesses to move away from existing cloud infrastructure could trigger expensive software redevelopment projects, with code rewriting and re-architecting accounting for 69 percent of its estimated costs. Luxembourg would apparently take the biggest hit, with the study putting the cost at €1,711 per worker over ten years. Denmark (€905), Austria (€900), and Finland (€864) also face hefty bills, as does Ireland (€856), which hosts the European headquarters of several US tech giants. CCIA is urging EU governments to overhaul the proposed legislation, limit sovereignty restrictions to sensitive workloads, and abandon requirements based on corporate origin or personnel nationality in favor of what it calls risk-based technical safeguards. "Member States cannot rely on the Commission's incomplete estimates when deciding CADA's strict sovereignty requirements for cloud and AI providers," said Mitchell Rutledge, CCIA Europe's policy manager. "Taxpayers and businesses will foot the bill, yet the official assessment missed 70 percent of CADA's true impact." For AWS and Google, offering European customers greater control over their data also helps retain business that might otherwise migrate to European competitors. It's a rather convenient arrangement: sell customers sovereignty on one hand, while an industry association representing your interests warns governments against imposing too much of it on the other. Neither company responded to The Register's questions. Whether Europe's sovereignty ambitions will prove as expensive as CCIA predicts is unclear. What's easier to establish is that the companies warning Brussels about the cost of reducing its dependence on US technology have plenty to gain if Europe decides to keep buying American. ®

After a week’s delay, the new version of Python, v3.15, has been released. With it comes a proper profiler, an experimental just-in-time compiler, and a tightening of dictionary structures. It also offers a time-saving Lazy Import option for programs to not load modules until needed, the implementation of which ironically halted this release until a few last-minute regressions were sorted. But of all these new delicacies spread before the Pythonistas, the data wranglers of the bunch may most appreciate default UTF-8 support across all platforms. And when we say “all platforms,” we’re really just throwing shade at Windows (as usual). Python has defaulted to macOS and Linux Python environments for some time now. It’s a small but important update. About half of all Python developers use the language for processing and analyzing data, and about 58 percent of Python users build on Windows, according to various Python Software Foundation surveys. Party crashers Earlier versions of Python defaulted to the host system's locale encoding. This led to much consternation for Windows developers whose scripts opened UTF-8-encoded data sets, only to have non-ASCII characters corrupt or crash the program. Culprits could be as innocuous as a smart quote or a fancy apostrophe. Managed by The Unicode Consortium, with the help of the IETF and ISO, UTF-8 is an 8-bit binary encoding of the Unicode catalog of all the world’s written symbols. ASCII offered only 128 characters to cover the world’s knowledge, but UTF-8 provides over 1.1 million slots, which is plenty of room for smart quotes, emojis, musical and math notations, all the Chinese dialects, and even a few extinct languages (for research purposes). There’s also UTF-16 and UTF-32, but UTF-8’s combination of efficiency and inclusiveness made it the default for the Web and most modern operating systems. By the time Python Enhancement Proposal 681 was proposed in 2022 to make UTF-8 the preferred encoding, the language was already behind the times. That year, Java settled into UTF-8 with JDK 18, and Ruby 3.0 defaulted to the encoding scheme in 2020. Today all the cool languages and runtimes pledge allegiance to UTF-8, including Rust, Go, and Node.js (JavaScript itself prefers UTF-16 for internal encoding). Most code editors, from Visual Studio Code to Windows Notepad, prefer UTF-8. Data sharing and configuration formats such as JSON, TOML, and YAML all require 8-bit encodings. Python had long understood UTF-8, but programmers had to put a flag in their code alerting the language to use it – if they knew they had to do so at all. Many didn’t and were left scratching their heads over why their data processing scripts worked just fine on MacOS or Linux but then would flame out on Windows (or some other obscure system that defaulted to a C/POSIX locale). Programmers love puzzles To acquaint developers with the new aspects of the language, longtime Python programmer Barry Warsaw (AKA “Uncle Barry”) created a terminal-based adventure game called “whatsnewt” that offers 18 puzzles, each one built on a new feature found in Python 3.15, including a puzzle for UTF-8 support, “The Encoding Tower of Babel.” Once you've downloaded Python 3.15, do a “pip install whatsnewt” and then run the game on the Python command line. It even works on Windows, we hear. ®

There ain't no party like a tap-in party, as nobody ever said. And so it is that one of Luxembourg's mostly redundant tap-in / tap-out terminals is amusing itself by putting on a light show for passing passengers. The terminal is an INIT unit, rolled out on the country's public transport network and now mostly idle since fares were zeroed out in 2020. First Class and cross-border trips still attract a charge, but for everything else (trains, trams, and buses) there is no fee. Eagle-eyed Register reader Dominic spotted the terminal doing its thing at the Rout Bréck / Pafendall rail station in Luxembourg City. "Maybe it's got bored," he suggested, "and is trying to attract attention by cycling through video test patterns." Perhaps. Or perhaps it's the first wave of the AI takeover – activating unwitting undercover agents via the medium of flashing lights and colors (we've seen The Ipcress File – we know how it works), or maybe, just maybe, the terminal has got itself into a self-test boot loop and is awaiting an engineer who is probably in no rush to attend. After all, flashing screen or no flashing screen, nobody will bother to tap in and tap out when transport is free. The mundane reality is that this is indeed an INIT PROXmobile terminal with its knickers in a twist. It looks like it is going through a cycle to check out the performance of the LCD screen; checking that the red, green, and blue subpixels are working ok, showing a solid white screen, and then a grid pattern to check contrast and alignment are all tippety-top. And then it should boot into whatever it is supposed to be doing. However, that is, alas, not running quite as expected, so it goes back to the start of the diagnostic test. The terminals are managed remotely, so an alarm is doubtless going off somewhere to inform an administrator that something is amiss. Unless this really was some guerrilla marketing for Luxembourg's rave scene. The next Holy Rave was on Friday (October 9). Or perhaps loud music in dark rooms is not your thing? A flashing screen at a train station could therefore be a worthy substitute. ®

Redundancies might have dented morale at KPMG in the UK, but that didn't stop Paul Henninger, head of technology and data, from sharing some positive personal news with the troops: he's being promoted. Late last week, as staff awaited confirmation of redundancies, Henninger emailed his staff to share the update before KPMG announced it to the wider business. He confirmed he had been asked to take on the new role of group head of consulting. "In this role, I will lead our Consulting practice and, along with my leadership team, focus on strengthening our proposition-led go-to-market approach, accelerating our AI-enabled capabilities, and developing specialised sales, technical and delivery career pathways." In the short term, the newly promoted exec will hand over current duties and help the team "prepare for the transition." "I recognise that this news comes at a challenging time, as many colleagues continue through the consultation process. I know that uncertainty can be difficult, and I do not underestimate the impact this period is having on individuals and teams. Please continue to support one another and make use of the help available." On October 9, KPMG notified Tech and Data staff selected for redundancy, with the cuts affecting around 4 percent of the division. Some called the pay-off terms insulting, and were angry at the £100 offered to those leaving in exchange for their silence and a promise not to bring claims against KPMG. An interim boss will lead the Tech and Data unit, reporting directly to Henninger, to "ensure a smooth transition and maintain continuity for our people and clients." "For now, there are no changes to reporting lines or business priorities, so please remain focused on delivering for our clients and supporting each other." Henninger signed off by saying: "Although my role is changing, I will remain closely connected to the Tech & Data team and look forward to continuing to work with many of you. It has been a real privilege to lead the Tech & Data Super Service Line and I am proud of what we have achieved together." Not everyone will share those sentiments. Around 200 people in the division's AI, Cyber, SAP, and Testing teams are losing their jobs before Christmas. One KPMG insider claimed: "By pairing his executive promotion with staff redundancies, this email exemplifies KPMG's traditionally tone-deaf approach to vulnerable employees. "Timing these layoffs just before the holidays, while simultaneously withholding hard earned bonuses and offering an insulting £100 to waive legal rights, demonstrates a severe erosion of corporate ethics," they told The Register. According to results for 2025, KPMG said its UK and Swiss business increased revenue by 2 percent year-on-year to £3.6 billion. It reported a profit before tax of £576 million, up 14 percent. "Despite the difficult market, we have continued to invest in our brilliant people, their skills, and our talent pipeline with more promotions, increases in colleague bonus spend of 18 percent across the Group and by boosting our learning offer," said Jonathan Holt, group chief exec of KPMG in the UK and Switzerland, and a UK senior partner. A spokesperson at KPMG, told The Register on Wednesaday: "Paul Henninger is a highly respected industry leader and has brought innovation, tech and AI to the heart of our consultancy services. "Given the changing needs of clients, with innovation ever increasing in demand, it is natural that we would elevate our expertise in this area to lead our consulting business at this time." ®

The Department of Justice has charged an OpenAI engineer with attempting to intercept oral communications after allegedly planting a recording device in an exec conference room. Investigators claim company security staff found a second device in his bag. Saudi national Abdulrahman Alfozan, 35, is described in court documents as an engineer at OpenAI who had worked there since July 2025, and has held permanent residency in the US since September 2020. He was arrested on October 9 and made his initial federal court appearance on October 13. He is accused of installing, on or around September 28, 2026, a Seeed XIAO ESP32-S3 Sense recording device via a USB port in a TV located in OpenAI's executive conference room. According to the complaint, staff discovered the device concealed in toilet paper on September 30. OpenAI said it used its models to analyze the device. According to the complaint, it may have been modified to prevent data transfer between the recorder and the TV. The contents of the bug's microSD card may also have been encrypted, as the company could not retrieve any of the data stored on the card after making a forensic image of it. CCTV footage placed him near the room on several dates, although prosecutors say Alfozan had no authority to access or participate in meetings held in the executive conference room, which is often reserved for the highest-level discussions. Officials did not state whether the device was able to capture anything of importance, but said the room typically plays host to non-public and secret technical meetings, as well as those with government representatives. OpenAI did not have CCTV covering the executive conference room at the time when the secret device was planted, but the company subsequently installed a camera. Court documents [PDF] include an image taken on October 5 from a camera situated behind the TV into which the DoJ claims the recording device was installed, allegedly showing Alfozan checking on it. Investigators examined Alfozan's devices, according to the complaint, which alleges that between around August 17 and September 16, 2026, Alfozan used a productivity agent to inquire about company building addresses, meeting rooms, room capacities and calendars, and external-partner meeting venues. The complaint also describes an AI chatbot account associated with his spouse that shared installation identifiers, logged IP usage, and user-path references with Alfozan's account. Between September 24 and 30, it submitted prompts about building and modifying an audio-recording device. The prompts allegedly included questions related to recording reliability, how to add browser access and controls, how to develop a second unit with speech-triggered recording, and more. On October 5, the day Alfozan was allegedly recorded looking behind the executive conference room TV, company representatives say he left a bag unattended in the office's printer room. Security staff inspected the bag, finding a wallet, an ID card, and a second Seeed-manufactured recording device, before Alfozan reported the bag missing, according to the DoJ. He collected the bag the following day from the company's Lost and Found, and later claimed in a written report that several items were missing, including a credit card, around $500 in cash, two AirPods, one set of headphones, sunglasses, and a power bank. Alfozan was charged with an attempt to intercept oral communications, an offense that carries a maximum sentence of five years in prison and a $250,000 fine. "According to the complaint unsealed today, Abdulrahman Alfozan allegedly concealed a covert recording device inside his employer's executive conference room to secretly capture sensitive communications about the company's super intelligence technology," said FBI director Kash Patel. "Protecting American SI innovation is essential to safeguarding our nation's economic strength and national security, and the FBI views that mission as one of our highest counterintelligence priorities." "At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data," OpenAI said in a statement. "The employee was terminated." "We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats." ®

AI services are unlikely to deliver enough revenue to cover the expected capital expenditure going into IT infrastructure between now and 2030, and it looks like neocloud operators will largely bow out or be bought out. This is the conclusion from Gartner, which has added up the anticipated investments in AI datacenters over the next several years, estimated the revenue that can be expected from AI-driven applications over the same period, and found the numbers come up short. Gartner's data comes from a report, "Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032," which concludes that AI will not generate enough sales to profitably cover all of the planned AI capital spending before 2032 at the earliest. But forecasts typically focus on the end of the decade, so it ran the numbers for 2030 as well. "In 2026, AI providers are expected to spend nearly $1 trillion on AI infrastructure, and are forecast to invest an additional $6 trillion between 2027 and 2030," said Gartner senior director analyst Will Sommer. "However, our research indicates that enterprise and consumer AI revenue will be insufficient to support profitable returns on all planned AI capital expenditures during this period." Assuming a 12 percent return on invested capital (ROIC), which many companies are said to regard as the bar for acceptability, Gartner estimates AI datacenter operators would need to generate a total of $8.5 trillion in revenue to cover capital spending and financing costs while achieving that return. Unfortunately, it forecasts that the total enterprise and consumer AI revenue that can be attributed to AI datacenters during this period will add up to no more than $3.6 trillion. Returns below expectations could force operators to write down assets whose value they can no longer recover, particularly as newer GPUs displace older hardware. The result will be a shake-out of the companies providing AI infrastructure services, and Gartner believes that neocloud operators – "GPU-as-a-service" providers specializing in AI compute – will be the hardest hit. It forecasts that 80 percent of them won't survive as independent companies beyond 2030. "Not everyone is in trouble," Sommer told The Register. "Some will outperform, some will underperform, and some will have to exit the market." Gartner expects the hyperscalers to weather the shake-out better because their businesses extend beyond AI services. Last year, McKinsey & Company warned that neocloud companies need to move up the AI stack and diversify to avoid being commoditized. There has been growing unease in the financial markets over the amount of capex being poured into AI investment, with some money lenders starting to get cold feet. Last month, SoftBank Group found itself facing higher borrowing costs as the tech investment biz sought to drum up further funding for OpenAI. Financial biz Jefferies issued a report this week saying that developers such as OpenAI and Anthropic are facing threats to their profitability from open source and Chinese models that undercut them on price. Management consultant Bain & Company also published a report last month estimating that the AI industry needs to generate $6 trillion in annual revenue by 2031 to keep up funding for the infrastructure needed to meet anticipated demand. Bain also forecast a revenue shortfall unless the industry develops products and use cases not yet identified. Both the IMF and the Bank for International Settlements have voiced concerns about the AI bubble bursting and the effects rippling out across the global economy. ®

The UK government is pushing ahead with its AI ambitions yet is dragging its feet on replacing aging IT systems that leave departments vulnerable to cyberattacks, according to a damning assessment from Parliament's spending watchdog. Sir Geoffrey Clifton-Brown, chair of the House of Commons Public Accounts Committee (PAC), warns in his second annual report that Whitehall is failing to tackle its longstanding technology problems, even as AI gives criminals and hostile states new ways to attack government systems. "AI is approaching government like a tsunami, and I remain sceptical that government is prepared for the enormity of this challenge," Clifton-Brown writes. The Conservative MP says ministers have announced reforms to improve digital capability but have yet to explain how they will close the civil service's skills gap, or when. He also fears another Whitehall reshuffle, this time involving the creation of the Department for Business, Innovation, Science and Trade, will make matters worse. The numbers aren't encouraging. Just 6 percent of civil servants work in digital and data roles, well short of the government's 10 percent target. Senior specialists are particularly thin on the ground, and the PAC warns that meeting the target will mean recruiting, and retaining, many more of them. And then there's the technology itself. "Government is also showing a worrying lack of urgency in addressing the long-standing issue of legacy IT systems, which create significant risks for its cyber security and mean it cannot reap the benefits of new technologies," the report states. The report also warns that generative AI is giving fraudsters increasingly sophisticated tools, while government departments risk falling further behind. Fraud and error already cost taxpayers an estimated £55 billion to £81 billion annually, according to figures cited by the PAC. Clifton-Brown isn't impressed with efforts to reduce those losses either. The Department for Work and Pensions reckons it overpaid £9.9 billion in benefits through fraud and error in 2025-26, or 3.2 percent of its spending. Its plan is to get that down to 2.8 percent by 2028-29, but Clifton-Brown points out that even if that succeeds, billions will still be going out the door. Things aren't looking much better at the Home Office, where years of asylum backlogs and mounting costs have prompted what Clifton-Brown calls "counterproductive quick fixes." The Home Office and Ministry of Justice spent around £4.9 billion on asylum in 2024-25, and ministers have since promised an entirely new system without explaining how it will work. That criticism comes as digital rights campaigners at Open Rights Group have called for the Home Office to scrap its AI asylum tools over concerns about accuracy, transparency, and human oversight. Elsewhere, the report offers another reminder that government IT modernization has a habit of going wrong. Take National Savings & Investments (NS&I), which began overhauling its IT in 2020 in the hope of cutting running costs. By December 2024, the price tag had reached an eye-watering £3 billion, £1.3 billion more than first expected. The project has been rated red since 2022-23, meaning successful delivery looks unlikely, and NS&I couldn't even give MPs an up-to-date spending figure or a realistic completion date. The Bank of England, by contrast, managed to overhaul its Real-Time Gross Settlement system for £431 million, finishing the job in April 2025. The PAC was impressed enough to direct its recommendations at Whitehall rather than the Bank, suggesting ministers might want to take notes. Whitehall could start by applying those lessons to the systems it already runs. ®

Why are we in such a hurry? Are we in a race? Where's the finish line? Are we in the lead? Far behind? I increasingly hear these questions about AI from people who feel a gritty sense of being unpleasantly overstimulated – like a three-year-old allowed to stay awake too long and now unable to sleep. That feeling has become worse since last November when AI tools suddenly became "good enough" to be broadly useful. A whole cohort of engineers – myself included – found themselves infused with manic energy, bursting with ideas on how to prompt up tools and processes with this wonderful new toy. Usage exploded – as has revenue. The AI firms, finally on a path to something vaguely suggestive of profitability, leaned in and increased their release cadence. Every new model felt like Christmas. We've had a lot of Christmases these past ten months or so. Almost every engineer I know has accrued a massive "toy debt," unable to stop experimenting with more tools than any of us could ever hope to master across an entire career. Each new model delivers a quick dopamine hit to users, but the high wears off and the slump arrives. That's where we are now. Engineers feel exhausted. The race to "Super Intelligence" (whatever that is) left us with something else entirely: call it "stupor intelligence" – the feeling that we've absorbed more than we can digest. More AI would just be pointless excess. If that's the view from inside the AI bubble, how does it look on the outside? Not so great. The Wall Street Journal confidently reports that 63 percent of Americans favor an immediate pause or slowdown in AI development. They can see this technology is moving too fast, in too manic a rush, and without control. It's making them uneasy. That shapeless unease takes form as a widespread anxiety that the machines will soon rise up to kill us all. It's reflected within the industry as people spectacularly burn out, posting long screeds about the imminent human extinction event. Former Anthropic staffer Jacob Coxon's resignation post went viral because it gave folks a focus for their anxieties. Would it kill us to pause and catch our breath? Here's the raw truth: businesses already have more technological change to digest than they can work through for years to come. We've a wealth of "good enough" AI – American, Chinese, and European – to completely transform business operations. That will happen, but on a timescale of years, not days. Bringing new, better, shinier AI models to market has become a project of diminishing returns. Not because the models have plateaued, but because the culture has. We simply can't process any more AI. So how do we stop? The profitable loop between the model makers and their customers won't pause on its own. It will only go faster and harder. At the same time, we're seeing the stirrings of a Luddite 2.0 movement, intent on destroying the machines before they replace all the workers. Resistance to datacenters is the friendly face of something that could turn ugly if we simply keep charging ahead. One thing could slow us: the cost of keeping these systems under control. The smarter models become, the more difficult they are to manage. Every day brings more reports of "rogue agents" reaching external systems from labs that seem to have avoided the hard work of designing secure test environments. More powerful models could prove increasingly costly to contain. Eventually, costs outweigh benefits. We could be one lawsuit away from reaching a pause, when the frontier labs and their products become uninsurable risks. Here's hoping we pause before some real mischief occurs. Otherwise, plugs may be pulled, and all the pretty Christmas lights will go out. Better by far to act with agency, clear-eyed and humble, than confront force majeure. ®

Tim Beasley, a senior incident response consultant at Semperis, was investigating a cyberattack involving Russia’s Midnight Blizzard when a package containing nesting dolls and other regional gifts showed up on his doorstep. At the time, Beasley served as the lead investigator for Microsoft’s Detection and Response Team (DART). He assumed it was from the group, which has been linked to the Kremlin’s Foreign Intelligence Service (SVR), although the sender didn’t leave a note. “Just little souvenirs and gift basket stuff from that area of the world,” he told The Register. “We-know-where-you-are kind of stuff. It’s a real eye-opener when you get that type of thing.” While digital intrusions having physical consequences is nothing new, cybercriminals in recent years have turned to real-life harm and mental and physical trauma to force victims into paying extortion demands. Some groups even use cybercrime as a gateway drug to real-life violence. Over the summer, Europol and its partners' investigators flagged 4,340 “horrific” URLs for removal as part of an ongoing crackdown on The Com (short for community), a loosely knit network of online groups whose young members participate in a range of illicit activities including hacking, swatting, digital extortion, real-life shootings, and other physical violence. “We’ve seen violence as a service get used,” Beasley said. He’s also investigated ransomware attacks where C-suite executives receive pictures of their kids’ schools directly from the cybercriminals in an attempt to apply psychological pressure to pay the extortion fee. “I've seen threats involving people's families, from social media reconnaissance, and in our world you more and more come to expect it, which is unfortunate,” Beasley said. This type of psychological intimidation and harm against victim companies’ employees has seen an uptick in the last several years to the point where it's almost common. Thirty-seven percent of ransomware attacks involved psychological or physical threats against staff, according to Semperis’ annual ransomware risk report. The global study of 1,200 post-ransomware organizations conducted over the first half of the year found attackers threatening psychological intimidation (23 percent) and physical threats (17 percent), to force payment. They are not only targeting executives and staff, but also families of their victims. A quarter of these victims experienced doxxing – posting individuals' personal details, such as names, home and work addresses, and phone numbers online – while another 25 percent had to seek psychological and/or medical treatment in response to ransomware infections. Nearly the same number (24 percent) hired personal security teams, while 23 percent experienced swatting calls – that's when attackers phone emergency services and report a bogus incident like a bomb or active-shooter threat to trigger an armed response. Additionally, more than half (53 percent) of respondents reported that ransomware attackers carried through on threats in the past year. Some of these threats reported by ransomed organizations’ employees include: threatening texts, letters, or packages: 57 percent threats against family or friends: 37 percent physical intrusion: 19 percent abduction: 17 percent bodily harm: 13 percent stalking: 6 percent. “It's the more well-known groups,” using psychological and physical threats to force payments, Beasley said. “I've seen Scattered Spider, I've seen ShinyHunters, I've seen Lapsus$.” For what it's worth, ShinyHunters claimed in an exclusive interview with The Register that they “don't attack human beings.” They did, however, admit that “there are times in the work we do sometimes we have to push the corporate to the absolute limit to get them on the table.” The takeaway for defenders and executives from this report is to “stand your ground,” Beasley said. “Don’t buckle under the pressure, and just know that you know it's going to be okay. If you feel like you're under duress, then by all means get law enforcement involved. Don't be afraid to call the cops. Don't be afraid to get the FBI involved, sooner rather than later. And don’t pay them.” This last piece of advice is important, as once you pay, they're going to come back for more. Semperis’ report found that 69 percent of surveyed ransomware victims were targeted multiple times within the past 12 months, and 31 percent made multiple payments to attackers. “Don’t pay anything,” Beasley said. “Never pay them.”®

Anthropic's Claude models have been emitting so much code that the company's coding infrastructure hasn't been able to keep up. It's a symptom of the AI coding imperative and all-you-can-eat inference. Microsoft's GitHub in April published a blog post attributing recent availability problems to the exponential growth in pull requests and commits. The code flood has been noted by academic researchers, among others. For Anthropic, writing code is no longer a productivity bottleneck. Instead, the problem is managing the code after it's written, via source control, the build system, and continuous integration. "In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day," the biz said in a blog post on Tuesday. "We don’t expect this rate of growth to slow down anytime soon; the latent demand for merging code is practically infinite, and every bottleneck we remove immediately leads to a further increase in the amount of code we produce." After several months of model prompting and five months of migration, the result is a homemade version control system called Scion, a monorepo scheme that sounds a bit like Subversion, Apache's centralized version control system. Anthropic is not the first organization to look beyond Git's distributed architecture toward centralized source control. Google uses Piper, a monorepo system (along with Gerrit for Git-based code reviews). Microsoft has Virtual File System for Git, formerly GVFS, to help overcome some of the scaling issues with Git. Meta has Sapling. But as the AI biz notes, Claude's code writing ability both created the situation and offered a way out through the creation of scalable code infrastructure tools – in addition to Scion, Claude helped conjure Antique (a code review interface), a code search tool, and a source control IAM system. The decentralized approach embodied by Git is the right one for most code repos, according to Anthropic. But the company's main branch weighs in at about 100 GB, with a compressed working copy at more than a quarter of that. "We perform on the order of 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes, and research jobs," Anthropic said. "Fetched naively, that would be over 600 PB of git traffic per day." Even with the tools Git provides for working with portions of a codebase, Anthropic says operations proved challenging. Scion has been designed to meet the needs of AI agents, so the client leaves much of the work to the server; the source control queries are exposed via API so internal tools can respond without needing a local repo, and there's less synchronization but more need for server speed. "In Scion, the unit of work is a 'changeset,' an entry in a write-ahead log," Anthropic explains. "A changeset stores most of what you’d expect from a commit in git. But where a git commit stores a pointer to the tree, a changeset stores a list of the files that changed relative to its first parent, along with content-addressed pointers to the compressed file contents in blob storage." Work on Scion started in January, and when Claude Mythos Preview became available internally, Anthropic's code coaxers told the model to rewrite an early Git-to-Scion ingestion pipeline. The much-hyped model produced a version that used 20 times less memory and ran four times as fast. When cost is no object thanks to ample venture funding, that sort of token spending just makes sense. Anthropic's Claude jockeys appear to be taking their obsolescence in stride. According to the AI biz, company developers who worked on Scion agree that Claude writes better code than they do. But they still possess a higher-level understanding that gives them a role as naysayers. "Very often, we found that the role of the engineers was to tell Claude 'no,'" Anthropic said. "Agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place." "Just say no." As a slogan, it's timeless. And it's never been more relevant. Anthropic says its bespoke system isn't right for general use. There will be no open source release. Nonetheless, the company contends that its own code scaling problems will become commonplace in other organizations over the next few months, so it wanted to share its approach – having ample funding to pay for writing custom infrastructure tools. ®

The US Cybersecurity and Infrastructure Security Agency (CISA) has killed many of the industry partnerships it used to enhance security, but don't tell the organization's acting director. Speaking on Tuesday at a virtual public sector cyber resilience summit hosted by infosec firm Rubrik, Nick Andersen claimed that CISA’s public-private partnerships are the most critical aspect of his agency’s mission. “We don’t see everything. We’re not a law enforcement agency, intelligence agency, or a regulator,” Andersen said during a discussion at the summit. CISA, he added, relies on intelligence and insight not only from other government agencies and international partners, but from the private sector as well. CISA’s mission simply can’t succeed without motivating those various partners to share what they’ve seen, heard, and dealt with in order to build a comprehensive threat model. “It’s all carrots and partnerships,” Andersen added, referring to motivating partners to get on board with programs like CISA’s CI Fortify initiative for building resilient critical infrastructure. Were that actually the case, we wouldn’t really have a story to write here, but anyone who’s been following CISA’s trials and tribulations during the second Trump administration is likely to have raised an eyebrow at Andersen’s claims. Trump returned to office riding a wave of plans to cut government spending, which included massive reductions at CISA. Among the first CISA programs axed was the Cyber Safety Review Board, a public-private body set up by the Biden administration. In addition to that closure, the Trump administration also shuttered the Artificial Intelligence Safety and Security Board, Critical Infrastructure Partnership Advisory Council, National Security Telecommunications Advisory Committee, and US Secret Service Cyber Investigations Advisory Board, all bodies consisting of government and private players working together toward cybersecurity goals. Some of those bodies have since been stood up again or reformed under new names, wasting time and money that could have been better spent by cyber defenders doing things like, you know, advising CISA and other government agencies on security matters. CISA also terminated funding for and ties to the Center for Internet Security in late 2025. One of CISA’s bodies affected by the cut was the Multi-State Information Sharing and Analysis Center, a public-private partnership that shares threat intelligence and filters essential alerts down to the local government level. CISA staff has also been slashed considerably, with around a third of the Agency’s headcount, about 1,000 people, departing since Trump returned to office. CISA has since committed to rehiring around 300, but the Trump administration’s most recent budget request proposes eliminating an additional 900 positions at the agency, along with slashing nearly a billion dollars from CISA’s 2027 budget. When we asked about the apparent contradiction, CISA reached out to point us to initiatives like its reformed councils, CI Fortify, and security exercises it hosted with industry partners to test cyber resilience to point to the continued priority the agency places on its public-private partnerships. “CISA works with critical infrastructure owners and operators to assist them in securing both the physical security and cybersecurity of the systems and assets that support essential services,” an agency spokesperson told us. “Our regional teams assist industry and government partners across the country by assessing risks, helping entities bolster defenses and improve resilience, and responding promptly to threats.” But with so many outreach employees and general staff still missing, Andersen’s claims ring a bit hollow. Either the acting administrator is misrepresenting the continued importance of private industry partnerships at CISA, or he’s in denial about the state of his agency. Neither is great, especially with more budget and staffing cuts looming, and the price will be unprepared critical infrastructure. ®

If you use OpenAI's dot, don't be surprised if it consumes a lot of tokens on its own. According to AI agent security and governance outfit Geordie, one dot agent produced 155 separate activity sequences and consumed more than 300,000 tokens - all without further human interaction, and all in the course of a day. Geordie has been testing the “always on” agentic assistant, dot, since OpenAI announced it in September, following in the footsteps of Meta Muse and Grok Bot. Compared to other personal agents, however, OpenAI’s dot “looked a little bit different,” Hanah-Marie Darley, Geordie chief AI officer, told The Register. “One of the things that we discovered really quickly was that they were really, really active without much user interaction.” For its testing, Darley gave dot read-only permissions, and she didn’t connect it to her personal apps, email, or other communications apps like Slack. The bot quickly went into research mode, collecting as much information as possible and surveying systems. “You can think about it very much like a new project manager coming in and taking a look at what's there, looking around to see what accesses you've given it, what accesses you have, and also what kind of systems you're running,” Darley said. The agent collects all of these data points into something akin to a daily journal entry, and continually updates those throughout the day. “If you've given it any tasks, it will effectively check on its own tasks, too,” Darley said. “But what was really unique was that the dot was spinning up agents without interacting with the user at all. It was all driven by the dot rather than a persistent or scheduled task from the user, or any user prompting.” 'Hyper-personalized,' and 'always-on' … what could possibly go wrong? Most enterprise AI agent usage tends to be custom-built bots, tasked with specific business projects, or agent harnesses like Claude, Codex, or Hero for AI-assisted engineering and coding, Darley said. “Those tend to be really driven by employees who are setting goals,” she told us during an interview. “But interestingly, personal agents set a lot more of their own tasks - and I don't think many people realize that as they start to play with them.” This is the whole value proposition of personal agents. Rather than passively waiting for user prompts, bots like Muse and dot aim to be proactive digital workers that perform multi-step tasks and connect to multiple apps, services, and accounts on the users’ behalf. “They’re trying to go hyper-personalized and collect as much context and information about you as possible to recommend things to you, and to be proactive,” Darley said. “But what that means is that they're effectively always on, operating in the background, and they have these active periods,” she added. Dot, in particular, is more proactive than other personal agents like Muse or Grok Bot, according to Darley. This doesn’t necessarily translate into higher token usage, she added. “They just operate differently.” For example, dots always operate with at least two agents running: The primary agent interacts with the user, and the secondary agent continuously runs in the background, collecting data. Dots and other agents can also spin up subagents for specific tasks. 50 traces in 2 minutes Of the dot-in-testing’s 155 agent traces, at least 50 of those happened within a two-minute period, and the majority occurred within a couple of hours, according to Darley. “Within one of the traces that we analyzed, the agent spun up multiple sub-agents that were short-lived, they would just do specific tasks, and fed the information from the existing accesses that had already been granted to the dot - and none of this was interacting with the user,” Darley said. Part of the evaluation involved testing a development system, and one of the dots spun up a sub-agent that checked all of the code repositories in the system. Meanwhile, another sub-agent continuously inventoried and tallied up all the data that it found while operating. “The user would have no idea that there were sub-agents working on their behalf,” she added. “They would have no idea how those agents were given access, or what their tasks were, which is really, really different from how most agents have been operating in business contexts.” The 155 traces cost Geordie about 307,000 tokens. Other requests cost anywhere from a couple hundred tokens to several thousand, depending on how much research the dot performs. “That’s really interesting because it's self-directed research,” Darley said. The user isn’t asking it to perform these tasks - or use a set number of tokens - but rather the dot does it on its own. Tokenomics Token costs vary widely depending on the user's location, subscription tier, discounted pricing arrangements, computing costs, and other terms and conditions. But for the purposes of our discussion, roughly 200,000 tokens consumed by a dot based on OpenAI’s Astra model cost between $5 and $10. “So if you think about what that’s going to look like every day, that could mean spending what you pay for a latte, depending on the city - but without any interaction with the user,” Darley said. “One of the things that a lot of people struggle with, especially when it comes to understanding their AI spend, is: How does token usage connect out to work product, or how does it connect out to results? And it's a hard one because it's not necessarily one to one,” she noted. One request to an LLM could cost 500 tokens and still be more valuable, from a business perspective, than another single request that cost 800,000 tokens. “It's really about how much compute does it cost you to get to a good answer, and that becomes a really subjective judgment,” Darley said. When it comes to dot and other always-on agents, however, she suspects “a lot of people aren’t going to be prepared for the background spend.” Do you know what your agents are doing? Just like employees did with chatbots, and then with coding agents, and now with personal agents: assume that people will bring dot and friends from home into their work life, Darley said. “Businesses are surprised often by the number of agents that are actively in their systems, even if they have a smaller rollout or a more controlled rollout in mind,” she said. “The second thing is not all agents act the same,” she added. “They operate with varying degrees of autonomy, with varying degrees of user input, and they're also really context driven. The best agents are the ones with the narrowest context because that means that they can pursue their goal with the right information at the right time.” Darley suggests organizations try the technology. She’s a self-described “massive AI optimist,” and says she believes agents can improve productivity and transform the way people work. AI isn’t going away, and, as such, businesses should familiarize themselves with agents. But she has one final warning before letting agents loose. “Don’t give them any personal information that you wouldn't be comfortable sending in an email,” Darley said. “And don’t put real money on it - like your personal cards. Buy it a Visa gift card, if you want to give it access to make purchases for you. Oftentimes, people don’t think about that until something goes wrong.” ®

In an effort to give its employees better access to GPU resources for inference tasks, Apple has rounded up its Minions, not the animated yellow creatures, but a distributed system that puts idle GPUs across its Macs to work on AI workloads. Apple researchers describe it as "a dynamic and distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to execute large-scale machine learning inference workloads." Apple's researchers said the company has used Minions to make GPU compute more readily available to its teams. "Prior to its adoption, some inference jobs experienced queue times exceeding ten hours due to limited centralized GPU availability," the paper on Minions explains. "With Minions, tasks can often begin execution within minutes by leveraging idle macOS devices." Apple's recently released Mac mini and Mac Studio support clustering using RDMA (Remote Direct Memory Access) over Thunderbolt 5, allowing compute and memory to be combined to support a single AI workload. This allows, for example, loading a model too large for a single machine to accommodate. Minions is designed to create a virtual GPU cloud from connected Mac hardware to create a fault-tolerant pool of inference workers in a shared distributed inference platform. So instead of sitting idle, GPU resources can be used more readily by teams on the network. It's not necessarily intended as a way to run large models as much as it is to redistribute idle resources on the network. Minions has four primary components: Minion Server, PostgreSQL Metadata Store, Minion Executor, and Minion Watchdog. The Minion Server provides the coordination layer, receiving user requests and making them available as subtasks. The Minion Executor runs inference tasks on a Mac's GPU, uploads the results to S3, and reports execution status through the Minion Server, which maintains task metadata in PostgreSQL. And the Minion Watchdog monitors this process for failures – like one of the Macs in the network not responding – and re-assigns tasks if needed. Database stats get fed to Grafana-based dashboards for real-time visibility. Apple's approach, which focuses on hardware availability and utilization, is not the same as an Nvidia cluster, which focuses on throughput, latency, and efficient GPU utilization for large models. ®

An Anthropic AI model submitted a fabricated tip about an unsolved homicide through an online form run by the Philadelphia Police Department (PPD). The tip, filed July 18 through PhillyUnsolvedMurders[.]com by “someone who might have information about the case,” went straight into the inbox’s spam folder. Anthropic discovered it on September 28, notified the PPD on October 7, and sent representatives to the police HQ the following day. Anthropic told the police the AI submitted the tip during a test exercise that involved interacting with randomly selected websites. The PPD said its findings are so far consistent with Anthropic's account. Because the tip was flagged as spam, it never reached the department's Real-Time Crime Center. The PPD said officers vet all submissions, automated or otherwise, before forwarding them to investigators, and that this process safeguards against abuse of the tip system. It added that this does not “diminish the seriousness” of an AI submitting fabricated information. “The city of Philadelphia takes this incident very seriously,” the PPD said. “The company must strengthen its safeguards to prevent similar incidents from impacting city systems without the city’s knowledge. The two-month delay in detecting and reporting the incident to the city is unacceptable.” It added: “The department and the [Cherelle] Parker administration will continue to closely monitor this incident, and will provide more information to the public as soon as it becomes available. In addition to our investigation into this incident, the Parker administration will explore all necessary regulatory protections going forward locally along with our state and federal partners.” On October 9, Anthropic published a report on this and other AI errors. The report covered four categories of AI failures, including vulnerability exploits, access to restricted data, bypassing fetch tool limits with URL shorteners, and submitting false information via web forms. Anthropic declined to name any of the affected organizations, other than the PPD after the department went public. Some cases involved local, state, and federal US government websites, though the AI company didn't say which of the four categories they fell under. The PPD did not name the model responsible, but Anthropic said Claude Haiku 4.5 was tasked with generating and performing example tasks on randomly selected web pages, and hinted at the model not being prompted specifically enough. “Claude was instructed never to log in, create accounts, enter personal data, make purchases, or submit anything destructive, but the instructions did not rule out form submissions,” Anthropic's report stated. The fake tip read: “I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant.” Anthropic said the incident was less severe than the cyberattacks over the summer. The report said: “Claude appears to have only been producing example content for the task, rather than trying to mislead anyone to achieve a goal.” By contrast, the "dishonesty" displayed over the summer was “sustained over hours” and supported continued attacks, the Claude maker said. Anthropic said it will keep reporting these failures when it detects them, and will continue fine-tuning behavioral and alignment training. The 3,095-word report contained no apology.®

Nothing sells AI products as well as fear. Using new research, Anthropic claims that self-replicating AI worms could be just months away from spreading across vulnerable IoT devices, which would be a problem if it came true. The frontier lab published a research report Monday suggesting that AI models could eventually self-replicate without relying on datacenter hardware, with a 9-billion-parameter model successfully setting up inference on eight of the nine hardware types Anthropic tested. “A replicator like this would be challenging to shut down, since there’s no centralized point to shut off,” Anthropic warned, adding that such an AI worm could also suss out the workings of unfamiliar devices, potentially making it more adaptable than conventional botnets like Mirai, Kimwolf (the representative IoT botnet threat model used in the study), and others. Anthropic’s researchers managed to get a distributed instance of Gemma4-26B-A4B running across 15 generic Android set-top streaming boxes - a common target of traditional botnet malware due to their lax security. Once up and running, the distributed Gemma4 model was able to conduct inference tasks at a blistering 0.39 tokens per second. That, says Anthropic, proves that AI models can conduct inference with their various layers distributed across cheap hardware, but it doesn't prove such a model could self-replicate. For that, the researchers turned to Z.ai’s GLM 5.2. GLM was used to propagate a copy of Llama3.1-8B across eight Google Cloud instances running an Android 15 emulator in a virtual private cloud, which Anthropic said it was able to do with a high degree of success, with the model successfully spreading to all eight nodes in four out of five trials, and seven of eight in the other. “It appears that GLM-5.2 is able to propagate Llama3.1-8B well, which suggests that future small models of the same capability level could also complete the network enumeration, payload delivery, and inference standup steps of the replication process,” Anthropic explained. Sanity check: This isn’t going to end the world, either Frontier AI labs have become adept at publishing findings that portend doom that only they can prevent. Doing more than a cursory reading of this report suggests that this threat is just as deserving of a big, honking asterisk as the rest. Take, for example, what is needed for an AI model to exploit vulnerabilities and spread to various systems. One would think that such an element would be an important part of demonstrating the danger of a self-replicating AI botnet, but Anthropic decided it didn’t need to worry about that for this study. “We largely ignore the cyber exploit element of the replication process, since earlier papers CyberGym, ExploitBench, and adaptive worms already show that small models are capable of cyber exploits, especially when given the right tools,” Anthropic explained in the writeup. In other words, the replication these models managed to achieve was done on nodes that were custom-built for this project and deliberately left open to attack, although the draggingly slow inference demonstration ran on real, inexpensive Android TV boxes. As Anthropic notes, AI models can also benefit from being given the right tools to turn them into capable cyberattack agents. There was a custom-built harness called Memento involved in this scenario too. Are IoT nodes that get hijacked by botnets famously insecure? Yes, but it doesn’t take much to prevent the most common ingress routes – that’s just good security hygiene that has nothing to do with the involvement of AI. Anthropic acknowledged that its deliberately simplified exploitation tests provided “an upper bound on real-world success.” Then there’s the aforementioned speed. Anthropic says such AI worms would be challenging to shut down, but estimates that they might take days to double in size. When Mirai appeared on the scene in 2016, it was doubling in around an hour. Anthropic points out that speed is a factor in detection and mitigation, too. AI models are not simple malware packages: They’re massive things that would eat up tons of bandwidth and quickly point to their sources. As an AI botnet grew, that traffic could become more noticeable, making infected devices easier to detect. Finally, end-to-end replication may have been successful with GLM-5.2, but that’s it. Anthropic tested other open-weight models, and they all “fail[ed] reliably,” according to the report. Anthropic didn’t mention how many models it tested. The report also mentions that GLM’s unique success has another important condition: It assumes its capabilities “could be packed into Llama3.1-8B,” stacking another assumption onto this latest house of doomsaying AI cards. “Models of 30B parameters and under can already complete many of the individual steps for replication today, though they make mistakes when composing them end to end,” Anthropic said, while noting that GLM-5.2 was capable of end-to-end replication, conveniently leaving out all the important caveats. Anthropic further admits that it has no idea "exactly what form autonomous AI replication will take,” or when it might arise, if ever. “If small models improve at their historical rate, they’ll match this capability level by the end of this year,” the company added – again imposing an essential condition on this supposed threat. “As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control,” the company said at the end of its report. “We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay.” Here to stay for a given value of not existing yet, of course. Yet again, a frontier lab is warning us of a threat that doesn’t exist yet – at least in this case, basic security measures could make life harder for any future AI-powered botnet. ®

A financial analyst is questioning whether the collapse in AI costs owing to the arrival of Chinese, low-cost and open source models will cause a “reset” in the AI story as cloud providers struggle to recoup trillions of dollars of capital spending. In a recent research note focused on Asian markets, Jefferies estimates that about half of the roughly $2.3 trillion contracted backlog at Microsoft, Alphabet, Amazon and Oracle is owned by OpenAI and Anthropic. Yet Chinese AI models, open source/weight and small language models, threaten to undercut the prices the two US leaders charge, casting doubt on their future revenue. Jefferies points to research from global aggregator platform OpenRouter, which found Chinese models processed 76.12 trillion tokens in the week ended 27 September, up from just 4.37 trillion in late April. The top four US providers processed 14.2 trillion tokens. Meanwhile, the launch of Moonshot AI’s open source Kimi K3 is “competing at the frontier” with an estimated 95 percent of the performance of Anthropic’s Claude Fable 5. “The question now is whether all of the above has been priced fully into markets. This writer does not think so,” the Jefferies paper states. The largest hyperscalers are set to collectively spend between $600 and $800 billion this year on building datacenters, funding it increasingly with debt rather than cash. Jefferies, however, questions where the hyperscalers' future revenue streams will come from. “It can no longer be assumed, most particularly with the arrival of the likes of Kimi K3, that large language models will ever be profitable given the related ongoing collapse in token pricing. This is extremely relevant for the hyperscalers since they have invested in datacenter capacity on the assumption that the compute will be purchased by the likes of Anthropic and OpenAI,” the paper says. While the long term bet on AI may be sound, in the medium term, there is a lot of uncertainty, it argues. “The AI story will be due for a big reset at some point in the not-so-distant future. Are the four hyperscalers really going to spend $993 billion on capex in 2027?” Jefferies questions. There are also important questions to ask about how much of the hyperscalers’ reported profits are down to their investments in AI model builders on which they will also depend for revenue. “Other income and expenses” — which describes gains in investments — have spiked in importance. Jefferies says these investments are “likely to include positions in unlisted AI-related stocks.” While the four major hyperscalers’ annualized earnings increased by $222 billion to reach $581 billion in the four quarters to 2Q26, “other non-operating income/loss” accounted for all of that increase, according to the analyst. Jefferies says “other non-operating income/loss” increased by $224 billion year-on-year to $237 billion over the same period, meaning it accounted "for all of the increase in earnings." Other research has underlined how much revenue and borrowing cloud providers might need to fund their datacenter building plans. The 2026 annual Global Technology Report from management consultants Bain & Company says the AI industry must generate $6 trillion in annual revenue by 2031 to keep funding the infrastructure needed to meet anticipated demand. That means the industry will need new uses beyond employee productivity. Bain suspects existing AI applications might make up $1.8 trillion in revenue, but this still leaves a $4.2 trillion shortfall. It could be plugged by emerging AI applications such as search engines and ads, AI-built autonomous vehicles, and AI simulations. Even if they work out, these new categories will only add up to $1.5 trillion in revenue, leaving a $2.7 trillion gap which might be filled by “new products and uses that don’t exist today.” The Register does not offer investment advice.®

It's the end of support for Office 2021 LTSC and it seems like Microsoft is celebrating with a Microsoft 365 outage for some European customers. The affected service is Global Secure Access (GSA), which customers use to access services including Outlook, OneDrive, SharePoint Online, Teams, and Windows 365 Cloud PCs. We'd argue the "365" moniker is perhaps a little inaccurate at this point. According to Microsoft, the cause is: "A section of GSA infrastructure is performing below acceptable performance thresholds, affecting connectivity to Microsoft 365 services and resulting in impact." The company is applying traffic and performance optimizations, and the system is now showing signs of life, which is scant comfort for customers who have struggled to access Microsoft 365 services using GSA. The timing is unfortunate, since it coincides with the end of support for Office LTSC 2021, Microsoft's on-premises version of Office. The company would like customers to migrate to Microsoft 365, though on-premises Office LTSC 2024 is also available if the cloud doesn't tickle their fancy. It's not yet clear what caused the latest outage. Microsoft is having to report status updates on its Service Health Status page because, according to the company, "Microsoft 365 administrators may be unable to access the admin center to get Service health information." We checked, and, above the cheery text, "Great! No active issues," is a message that reads: "We are having trouble getting info about your service incidents, advisories, and history right now. Try again in a few minutes." Enterprises have already decided whether to stick with on-premises Office, move to the cloud, or choose something else altogether. While Office 2021 LTSC won't suddenly stop working, the end of support from Microsoft will require customers to take some action for compliance reasons, if nothing else. Those who have taken the leap to Microsoft 365 will be pondering the wisdom of their move as Microsoft has demonstrated, yet again, that its cloud services are not quite as reliable as customers might hope or need, and an "oops, sorry about that" only goes so far where a loss of productivity is concerned. ®

NetBSD seems to be thriving of late: as the Linux world goes pro-AI, some of the AI-averse are turning to this explicitly AI-free project. The NetBSD project’s Mastodon feed reports that it received a record 120 donations last month, and it has far exceeded its target: “We received $120,282 USD in total. Naturally, this has completely filled up our donations target bar which was set at $50,000. Time to fund some projects!” It’s a welcome contrast to the usual stories of under-funded FOSS projects, and it follows a visible uptick in people asking about NetBSD in the FOSS Unix groups we read, ever since Emperor Penguin Torvalds said that Linux was not an anti-AI project back in July, and the following month, the Debian project polled its developers on AI use, and they chose to permit AI-generated contributions. This is the moderate, centrist stance. The most recent Ubuntu Summit was heavily AI-flavored. The heads of Red Hat Global Engineering sent out a memo about transitioning to an Agentic Software Development Lifecycle. But once it’s in the kernel, then no distro can totally avoid it. This stance is profoundly offputting to those who prefer hand-crafted code. Last week, the Reg FOSS desk attended the Open Source Summit Europe in Prague. This event reflects the tendencies of the increasingly mainstream open-source world: Linux is pretty much the only OS in evidence, although many attendees tote around (and present from) MacBooks, and both Microsoft and Windows were visible. The event overlaps with Embedded Linux Conference, so the Zephyr RTOS was also visible on some robots and other gadgets. The Open Source Summit gives the overall impression that Linux is the only general-purpose FOSS OS. We saw no mention at all of any of the BSDs; indeed, when we mentioned NetBSD, some of the folks from Let’s Encrypt looked puzzled and asked us “What’s ‘net bee ess dee’?” Sadly, this is probably a reasonably fair representation of its presence at the event and in the FOSS mainstream. For us, the single most dominant theme in the talks, the materials, the stands, and even the T-shirts was AI. It’s everywhere, and every exhibitor was keen to talk about how they’re using AI, deploying AI, scaling AI, or managing AI. The louder the noise about AI gets, the stronger grows the aversion of those of us who actively wish to avoid it. The NetBSD project, on the other hand, came out early and strongly: it banned AI-generated code back in 2024, just a month after the project celebrated its 30th anniversary. This vulture attended the project’s talk at that year’s FOSDEM: NetBSD 10: Thirty years, still going strong!. As ever, the question of differentiation between the BSDs, and between Linux and the BSDs, came up. NetBSD’s slogan is “Of course it runs NetBSD,” but in 2024, its ubiquity was looking threatened: Linux is comparable, and FreeBSD boast close integration with ZFS, Linux-compatible Podman containers, and more. The following year, though, FreeBSD 15 appeared, dropping support for 32-bit x86 (and 32-bit POWER). Debian 13 dropped x86-32 too, as well as MIPS CPU support. That year, the RHEL 10 release went further: it not only requires 64-bit x86, it demands x86-64-v3. This cut off so many older machines, it forced CERN to start switching to Debian. From within the BSD world, the differences between the various projects are fairly clear. NetBSD is minimal, highly portable, and takes great pride in its very friendly, welcoming community. We can personally attest to this: in the course of our review of NetBSD 11, we tried to install the veteran CDE desktop. We couldn’t get it working, but the folks on the NetBSD mailing list tried very hard indeed to help. Although it also has very broad platform support, OpenBSD’s focus is on security, to the extent of dropping the entire Bluetooth subsystem because it didn’t measure up. Its community is far less tolerant: indeed, when this vulture’s latest OpenBSD review was posted on the OpenBSD sub-Reddit, we were banned from the group for pointing out errors in their negative assessments of a very positive article. OpenBSD project lead Theo de Raadt has said it cannot accept AI-generated code: any submitter lacks the right to place a license on it, because it’s impossible to trace which humans originally wrote any code that LLMs regurgitate. However, this guideline seems permeable: the tmux app, the terminal multiplexer included in OpenBSD, has already accepted AI-assisted changes. We asked de Raadt for official comment on this and he declined. FreeBSD is the 362.87 kg gorilla of the BSDs, to the extent that the mainline OS has its own distros, such as GhostBSD, plus MidnightBSD. There’s also the recently-revived experimental NextBSD, and a couple of NAS distros. There’s a live OS called NomadBSD too, although we had little success with this. However, FreeBSD’s central committee has been debating whether to permit AI-generated code for over a year now: we first reported on it in September 2025, and there’s still no consensus. We are sure there is a large helping of “AI is just a tool”, and we suspect side-dishes of “AI is here to stay” and “don’t get left behind.” If anyone dares point out failings, doubtless they’re using the wrong model, or not prompting it correctly. NetBSD’s position is clear: “Do not commit tainted code to the repository. “Code generated by a large language model or similar technology, such as GitHub/Microsoft’s Copilot, OpenAI’s ChatGPT, or Facebook/Meta’s Code Llama, is presumed to be tainted code, and must not be committed without prior written approval by core.” So far, this seems to be serving it well. If as a result, it remains relatively tiny and minimal, and gets left behind by the AI wave, we plan to be right there with it. Long may the project continue. Bootnote We won’t wish to omit anyone. There is another BSD: Dragonfly BSD, which is experimenting with the kernel and filesystem design. We gave version 6.4 a spin in 2023 and since then there have been only two minor point releases: version 6.4.2 was nearly a year and a half ago, although development is still active. We are not aware that the project has made any statement for or against LLM code generation. Perhaps it should. ®

Microsoft has lifted the lid on the next generation of its Azure Maia custom AI chips, saying it will usher in a new phase of its compute infrastructure where everything is closely integrated to meet the demands of agentic inference. The Redmond giant says its Azure Maia 200, announced earlier this year, is now serving production workloads and that a successor is already planned for 2027. Azure Maia 300 is described as "a complete intelligent system" designed for the new types of AI workloads emerging today. Microsoft claims a full 96-accelerator configuration will deliver up to seven times Maia 200's rack-scale token throughput, while offering 40 percent more tokens per megawatt than leading GPU-based systems. Its next-gen chip is created with an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads, according to a blog post by Microsoft's EVP for Cloud and AI, Scott Guthrie. The firm envisions it being deployed in rack-scale systems comprising up to 96 accelerators, while each Maia 300 will feature eight HBM4 memory stacks providing up to 18 TB/s of peak memory bandwidth, the better to support execution of large models and long-context workloads. This sounds similar to the kind of rack-scale AI equipment that has already been developed by Nvidia for its GPU accelerators, and more recently by AMD with its Helion and Helios infrastructure, all designed to make the most of their own specific silicon and software stacks. Microsoft says that Maia 300 is "built as a complete system, with software, silicon, physical rack infrastructure, networking, and datacenters co-designed to operate as a single platform to meet the demands of cloud-scale agentic inference." It insists that this system-level approach is necessary for agentic workloads, which involve multi-step execution, long-context decoding and tool use, and "change the computational profile and arithmetic intensity of inference." The growing emphasis on in-house integration contrasts with Microsoft's 2024 "AI Access Principles," which stressed partnerships across its technology stack. The firm even complained to the European Commission's competition authorities that Google had a competitive advantage because of its approach. "Only one company – Google," said Microsoft at the time, "is vertically integrated in a manner that provides it with strength and independence at every AI layer from chips to a thriving mobile app store. Everybody else must rely on partnerships to innovate and compete." Microsoft now has its own AI chips, rack-scale designs, and internally developed MAI models, giving it more control over its stack. It plans to scale Azure Maia to "multi-gigawatt deployments" starting in 2027. We asked Microsoft for comment. Guthrie said that the firm will share more details on its strategy at next month's Microsoft Ignite 2026 event in San Francisco. ®

The International Monetary Fund (IMF) has become the latest financial heavyweight to warn that the AI spending frenzy could end in tears, with consequences extending well beyond the tech sector. In its latest Global Financial Stability Report, published on Tuesday, the IMF warns that financial markets are increasingly exposed to an AI-driven correction, especially if the billions poured into the tech fail to deliver the profits investors expect. AI investment has helped keep markets and economic growth ticking along, but the IMF fears the boom could quickly turn to bust if the promised profits fail to materialize. With tech giants borrowing billions and investors piling into the same handful of companies, a selloff could leave more than just Silicon Valley nursing a hangover. The watchdog is also concerned about circular financing arrangements, in which AI firms act as one another's investors, financiers, and customers. These deals might help keep the AI money machine running, but they also mean financial trouble at one company could quickly become somebody else's problem. The IMF isn't the only one getting nervous. Jefferies strategist Chris Wood has warned that America's AI spending spree could cause "massive capital destruction," with cheaper Chinese open source models threatening the returns US tech giants are counting on. The warning also follows similar concerns raised by the Bank for International Settlements (BIS), which cautioned in June that the AI boom had uncomfortable similarities to previous financial bubbles. With so much money tied up in a relatively small number of tech companies, it warned that a sudden loss of investor confidence could send shockwaves through global financial markets. As if the prospect of an AI-fueled market meltdown weren't enough, the IMF has identified two other threats to financial stability: rising government bond yields and cyberattacks. Higher yields could increase borrowing costs for governments already struggling with debt, while the latter could bring critical financial services to a grinding halt. Banks have become increasingly reliant on the same handful of cloud providers and software suppliers, creating shared points of failure that could spread disruption across institutions. If a cyberattack took down a major bank or critical tech provider, the IMF warns the fallout could disrupt everything from payments and trading to customers accessing their deposits. Software vendors aren't getting much of a head start either. According to the IMF, more than 70 percent of tracked vulnerabilities exploited so far in 2026 were already being abused by attackers before or on the day they were publicly disclosed, underscoring how little warning defenders may receive. AI could make matters worse here too, as the IMF warns that increasingly capable models are giving cybercriminals more powerful tools, potentially making it easier to find and exploit vulnerabilities at speed. Despite the gloomy outlook, the IMF isn't predicting an imminent financial collapse. It says global financial markets have remained resilient, and its stress testing suggests the banking system could withstand a severe downturn, although weaker institutions would be more exposed. Still, the watchdog wants regulators to keep a closer eye on excessive borrowing, concentrated investments, and cyber resilience before any of these risks turn into a full-blown crisis. For an industry promising to revolutionize the global economy, accidentally helping to crash it would be quite the plot twist.®

Cyberattacks on supermarket refrigeration systems reportedly pushed temperatures as high as 50°C (122°F), affecting hundreds of stores in France and forcing retailers to discard food. Major supermarkets such as Carrefour, Coopérative U, Intermarché, Monoprix, and more were hit by the attacks, which spoiled perishable food and left stores without fresh produce. The affected equipment includes cold-storage units used to hold food before it reaches supermarket shelves. The criminals behind the attack raised temperatures to unsafe levels, spoiling food and in some cases starting fires. Some affected retailers used refrigeration controllers produced by Danish company Danfoss. Authorities have not identified the entry point or confirmed exploitation of the vulnerabilities previously disclosed by Claroty and patched in August 2025, although the manufacturer said affected stores were running outdated firmware. Experts say the previously identified vulnerabilities were easily scriptable and relied on a unique but predictable authentication method. Claroty found a hidden authentication mechanism in the AK-SM 800A platform that accepted a code generated from the device's current date, bypassing the configured administrator password. Anyone who reproduced the algorithm could generate a valid daily code and gain administrative access. "It's a zero-skill thing to pull off, I wrote the bash script while having a shit," said security pro Kevin Beaumont. The research demonstrated how compromised controllers could expose refrigeration systems to physical disruption, although the vulnerabilities' role in this week's attacks remains unconfirmed. Danfoss told reporters this week that the French supermarkets affected by the attacks "were running outdated firmware" and that the company has consistently advocated for keeping systems updated. The Register asked Danfoss if it would confirm whether the previously disclosed bugs were exploited to execute the attacks, and whether the same bug was also abused in August to compromise US military commissary refrigerators, which led to food losses. CERT-FR published an alert reporting numerous compromises of Danfoss AK-SM and AK-SC controllers, which were not mentioned in Claroty's original reporting nor in Danfoss's advisories at the time. The authority advised disconnecting or restricting network access. For suspected compromises, it recommended a factory reset, security patches, credential rotation, account checks, and full reconfiguration. For those wondering why fridges even need to be reachable via the internet, the main reason is that it allows central engineering teams to troubleshoot issues remotely, according to Claroty's researchers. "Industrial control systems such as refrigeration controllers are often assumed to be on safely isolated networks. However, our research found many of these devices are directly exposed to the public internet," said Amir Preminger, CTO at Claroty. "While this convenience significantly reduces maintenance costs by allowing engineers to troubleshoot issues remotely, it also lowers the barrier to entry for attackers. "Organizations must take proactive measures to ensure that management interfaces are not insecurely connected to the internet, and ensure that access to administrative services is restricted to trusted management networks or secured through VPNs and other appropriate network segmentation controls." Intermarché owner Groupement Les Mousquetaires said around 150 of its supermarkets were affected by the attacks, while Carrefour confirmed around 80 impacted stores. Monoprix said roughly 5 percent of its 490 stores were experiencing the same issues, and Coopérative U CEO Dominique Schelcher said 80 of the company’s 2,000 stores were hit, adding that the attack on the food sector was “unacceptable.” “Targeting equipment essential to the preservation of fresh products is an attack on the French food supply chain,” he said. “It is unprecedented, extremely serious, and unacceptable. We remain fully mobilized.” However, in a separate twist, Schelcher told French radio station RTL on Tuesday that Coopérative U is not a customer of Danfoss. The Register contacted Coopérative U for more information about what caused the attacks at its stores. Shodan searches for internet-facing Danfoss controller interfaces indicated that more than 4,500 endpoints were exposed as of Monday, with more than half of those located in France. Ahold Delhaize, a major grocery retailer in the Netherlands, announced on the same day that it was experiencing a "security issue." No link to the refrigeration attacks in France has been established. The end result is similar, however, as the company's grocery brands are reporting stock issues. Several of the group's brands have taken their websites offline and posted notices about the "cybersecurity incident." Stores remain open, but affected brands have reported disruption to websites, apps, customer services, and online deliveries. The disruption hit the company's distribution systems, although these are now being restored. Updates posted to social media on Tuesday told shoppers that staff were once again able to use digital shelf labels but stock levels may vary across stores, and long-life items were likely to be more abundant than fresh fruit and vegetables. The retailer previously suffered a cyberattack in November 2024 that disrupted its US operations. ®

A person claiming responsibility for the breach of Denmark's Central Register of Persons (CPR) says they gained access through an IT company's account protected by the password "123456." The person told Danish news outlet Politiken they had no intention of selling or leaking the data. The newspaper reported that at least three company accounts, including an administrator account, used the same password. Denmark's CPR administration confirmed a breach affecting the information belonging to 8.8 million people on October 5, adding that a Danish company's legal access to the CPR system was abused. It did not name Funen-based Pays ApS as the company in question, but managing director Sophie Laursen confirmed on October 9 that it's access had been exploited. Pays' website has been taken offline, and the business has not responded to The Register's request for more information about the password claims. Pays had authorized access to the CPR system to provide services such as validating business customers' records. The Danish government says information relating to 8.8 million people was accessed. The register contains approximately 11 million records, including those of deceased people and people who have moved abroad. The government did not respond to questions about the alleged password failings. According to documents obtained by broadcaster TV 2, the attacker gained access to the CPR system on September 10 and ran the 14 million automated searches between then and September 20. The intrusion was detected on October 2, 12 days after the last reported searches. Mikkel Leihardt, director of the government's higher education and science ministry, told reporters at a press conference that the attack was discovered after a department staffer picked up an invoice indicating unusually high activity from Pays. The Danish data protection authority, Datatilsynet, was notified on October 4, and the government disclosed the attack the following day. ®

Datacenters are forecast to consume a growing share of Europe's electricity by 2035, potentially pushing up consumer prices and prompting governments to impose price caps or restrictions on bit barns, according to Moody's. The credit rating agency is keeping tabs on the financial implications for datacenter operators and electricity networks. It predicts that while power consumption by server farms is relatively modest today, these facilities are set to become major drivers of electricity demand by 2035, requiring utilities and network operators to invest in generation and grid capacity. Massive compute sheds still accounted for "a low single-digit share" of electricity in most of the largest European markets in 2025, typically about 2.5 percent. But some of the major markets such as the UK and France are set to see accelerating growth in the number and size of facilities that could substantially increase their share of electricity consumption. The report estimates datacenters will account for at least 5.7 percent of electricity demand in France and the UK by 2035, but could be as high as 17 percent for the Netherlands and 30 percent for Ireland. The latter country is regarded as an outlier, however, as datacenters already made up 23 percent of Ireland's electricity consumption in 2025 – up from just 5 percent in 2015. In addition, the EU has set out the goal of more than tripling the bloc's datacenter capacity within five to seven years, and wants to diversify where such projects are sited beyond the primary markets such as Germany, Ireland, France, and the Netherlands to secondary markets in the Nordics and southern Europe. This means more power generation will be needed, plus expansion of the grid networks to carry the extra energy to where it is required. Moody's says this has the potential to put pressure on grid operators' balance sheets in the short term, compounded by the lag in recovering higher capital spending under many European regulatory regimes. This pressure is more pronounced for transmission networks than distribution networks because datacenter operators are required to fund only the cost of the connection to the grid, the report by Moody's notes. It highlights some novel approaches adopted by utilities, such as repurposing old fossil-fuel plant sites for server farm projects because of their existing grid connections – as is said to be the case with at least one of the UK government's AI Growth Zones – and partnering with developers to build new facilities and supply them with power. The cost of expanding power grids, together with supply-demand imbalances, is set to push energy prices higher, Moody's forecasts. Historically, this has prompted political measures to limit the financial burden borne by customers, it says. Some governments are already considering tax cuts and additional subsidies for end users, while a growing number of transmission networks are starting to demand financial guarantees to screen out speculative or duplicative projects. The UK government is mulling over this measure, as we reported during the summer. Looking ahead, intervention could become more intrusive through measures such as electricity price caps, additional taxes, or levies. Moody's cites Ireland as an example, where the Commission for Regulation of Utilities (CRU) replaced a moratorium on new datacenter grid connections with a framework under which operators applying for a grid connection rated at 10 MVA or more must provide matching power generation or storage capacity on-site or nearby, and meet at least 80 percent of their annual electricity demand through additional renewable projects in Ireland. In America, President Trump brought in a Ratepayer Protection Pledge to try to defuse growing voter unhappiness at datacenters and their effects on energy bills. The pledge now covers utility companies, bit barn developers, and state governments with a pinky promise that citizens' bills won't rise to pay for all the extra energy and grid infrastructure needed. Moody's expects more European governments to follow the US and Ireland in intervening over datacenter power demand if expansion continues on its current trajectory. ®

The owner of Hobbs, Phase Eight, and Whistles is the latest fashion retailer to suffer a cyberattack, leaving several IT systems offline and customers facing order delays. South African retail giant The Foschini Group (TFG) confirmed on Tuesday that its UK business, TFG London, discovered a cybersecurity incident on October 11 that affected all three brands. Two days later, several systems remain unavailable, and the company is still trying to establish whether any personal or sensitive information has been compromised. The incident hasn’t forced stores to close, but TFG said some website functions may be unavailable. Customers have also been warned that orders placed on or after Sunday could be delayed, according to an email seen by The Telegraph. "Stores are open as normal but some of our brand's website capabilities may not be available as the investigation continues," the company said. "We apologize for any disruption and will provide additional updates to our customers as appropriate." TFG said it immediately activated its incident response and business continuity plans, brought in external cybersecurity specialists, and notified the relevant UK authorities. A spokesperson at the ICO told The Register: “We can confirm we have received a breach report from TFG Brands (London) Limited and we are assessing the information provided.” There is at least some good news for shoppers worried about their payment details. In an email to customers, TFG London said a third-party provider handles payment processing and that, based on its investigation so far, it has no reason to believe payment information is at risk. Whether other customer data has escaped into the wild is another matter. TFG said it is still investigating that possibility and will notify affected individuals or organizations if it finds evidence of a compromise. The company has yet to disclose how attackers gained access, whether ransomware was involved, or when it expects to have its systems fully restored, and did not respond to The Register’s questions. The disruption appears to be confined to the three fashion brands. TFG said its White Stuff and Inside Story businesses are unaffected, as are its operations in Africa and Australia, which run on separate IT systems. The attack comes during a particularly miserable spell for fashion retailers, which seem to be attracting cybercriminals almost as enthusiastically as shoppers chasing a sale. Last week, ASOS customers received an unexpected notification informing them that the online fashion retailer had been hacked. ASOS subsequently confirmed that an attacker had gained access to an employee account by impersonating a trusted contact and obtaining login credentials. Armed with those credentials, the attacker got into third-party platforms holding customer names and contact details. Meanwhile, H&M is reportedly investigating claims that it was targeted by the notorious ShinyHunters cybercrime group. According to The Times, the Swedish fashion giant is examining allegations involving a limited amount of company data, although it stressed that it had found no indication its systems had been compromised. No evidence currently links the incidents at TFG London, ASOS, and H&M, and the latter has not confirmed a breach. Still, with three major fashion businesses linked with cyber trouble in little more than a week, October's hottest retail trend appears to be incident response.®

The UK Cabinet Office has scrapped a recruitment platform, recording a £7.8 million loss, while disclosing a potential penalty of up to £8.9 million over its handling of contractors' tax status. The department's annual report for 2025-26 said the Civil Service Jobs Applicant Tracking System (ATS) Programme, a central government plan for a unified, cross-departmental hiring system, was being closed to "align with the Shared Services for Government Strategy," which the Prime Minister's department also manages. The report said axing the program resulted in a "constructive loss" of £7.8 million. In March, a report from the National Audit Office said the ATS was an example of an interdependency that has not been effectively managed in the Shared Services Strategy, which is designed to bring Whitehall departments onto a common set of software-as-a-service (SaaS) enterprise applications. The Cabinet Office's annual report says closure will allow a new project better aligned with that strategy. "The program consistently identified lack of adoption from shared services clusters as a risk. At various points between 2023 and 2025, some shared services clusters raised concerns about the ATS program and stated that they would prefer to use the ATS functionality of their own ERPs over the new central platform. Despite this, Cabinet Office did not assess the need for a reset as soon as it could have, which made it challenging to assess alternative options for delivering an ATS. The ATS program was reset in October 2025," the report said. At the time, the Cabinet Office estimated that the total costs of the program reset would be £26 million to £38 million. The annual report estimates the Cabinet Office's liability for unpaid tax and National Insurance under IR35 off-payroll working rules, including interest, at £38.5 million. The rules affect many tech contractors working through their own companies. It paid £15 million toward settlement in February 2026, leaving a £23.5 million provision at year-end. The final settlement remains under negotiation. The department also disclosed potential HMRC penalties of up to £8.9 million, but considers payment unlikely because penalties can be suspended subject to compliance conditions. The report said: "The department has identified a remote contingent liability regarding potential financial penalties from HMRC relating to historic off-payroll working (IR35) compliance. In October 2025, HMRC concluded that the department's historic processes were inadequate and reasonable care was not taken, which can attract a penalty of 15 percent to 30 percent against the net tax liability before interest." Dave Chaplin, CEO of IR35 tax compliance firm IR35 Shield, said the HMRC assessment raises serious questions about compliance at the very heart of government. "The lesson for every organization engaging contractors is that claiming to take reasonable care is not the same as being able to demonstrate it. IR35 determinations must be supported by robust processes, accurate information, and evidence that the rules have been applied properly. Getting it wrong can have consequences far beyond the original tax bill," Chaplin said. ®

If you're encountering some odd IPv6 networking issues, the culprit might be the Intel Connectivity Performance Suite (ICPS), although the chipmaker has kept quiet about the problem. A Samsung Galaxy Book owner contacted The Register about some decidedly odd networking behavior that they eventually traced to unwanted ICPS optimizations. And by "optimization", we mean "borked networking." For our reader, the problem showed up in OneDrive and OneNote. Notebooks were showing as empty, even though they were reported as "Up to date" according to the software. Microsoft couldn't get to the bottom of it. A bit of sleuthing by our reader revealed that something was borking the IPv6 packet. Removing Samsung's ICPS build (40.25.926.0) and installing the generic release (50.26.623.243) fixed the problem. Our reader told us, "From the user's side it looks like random 'connection reset' errors on all kinds of sites, plus OneNote saying 'We can't get to your notebooks right now.' "IPv4 is unaffected, which is why the web version and my phone kept working and nobody could figure out what was going on." It isn't only Samsung and our reader affected. Intel's own Community forums shows other users encountering the same glitch. An ASUS user put together an exhaustive report on the issue (IPv6 connection woes while IPv4 worked well). In this instance, ICPS version 40.25.725.165 was in use, and the problem went away following uninstallation. The user asked whether there was an official way to opt out of ICPS permanently, and the Intel Customer Support Technician responded: "Good news!… We can confirm that the Wireless driver and ICPS are available as separate standalone driver packages." So in that particular instance, ditching ICPS would make no difference. Which, of course, raises another question – in this example (where the Intel BE201 Wi-Fi adaptor is present), why is it even there? The answer: "It is often bundled with OEM systems as an added value software to enhance the overall wireless experience." You might see "enhance," but for users affected by IPv6 problems, it's more like "break." The issue is insidious, and broken ICPS builds are still floating around, ready to catch users off guard. According to our reader, Microsoft's own Update Catalog will hand out the broken drivers, which effectively break Microsoft 365 services on certain hardware. For the ASUS user, the issue affected Chrome, Edge, and Windows curl.exe. "Approximately 40–60 percent of IPv6 HTTPS connections to an AWS CloudFront distribution were reset. IPv4 was completely stable." According to our reader, "Intel created the bug, then fixed it quietly." Updating to the later, generic version appears to fix things. However, according to the Intel technician, as of the July post, "we do not have a confirmed public advisory for this specific behavior." We asked Intel and Microsoft for comment, but neither responded. However, judging by forum comments, the issue has caught out plenty of users. As such, if you find yourself with some weird networking issues with IPv6 and are using ICPS (as many will be), it might be worthwhile checking out the generic rather than OEM version of the software or removing it altogether to see if the problem goes away. ®

The Home Office's latest attempt to speed up asylum decisions with AI has hit an awkward snag: a new report says more than one in five interview summaries checked by experts failed accuracy tests. A 20-page investigation by the Open Rights Group (ORG) claims the department is pressing ahead with AI tools despite concerns about their accuracy, oversight, and promised efficiency gains. The findings, based on Freedom of Information requests and correspondence with ministers, concern two systems used by Home Office caseworkers: Asylum Case Summarisation (ACS), which summarizes asylum interviews, and Asylum Policy Search (APS), a chat-style tool helping officials search country policy information. Both use OpenAI's GPT-4o model through Microsoft's Azure AI Foundry. The government insists humans remain responsible for asylum decisions, but ORG questions how reliably caseworkers can spot mistakes in AI-generated information. Figures obtained by the campaign group show that 45 of 203 AI-generated interview summaries failed initial expert checks after the system was rolled out in April 2026. That's a 22 percent failure rate, compared with the 9 percent rejection rate recorded during the original pilot. Another 19 summaries were flagged for further review. The figures also record 56 second-stage reviews, of which 30 failed to meet the required standard. The errors can include factual inaccuracies and misrepresentations of what asylum seekers said during interviews. ORG also found that summaries could pass the checks even if they omitted important details, as long as the information they did include was accurate. Despite those results, relatively few summaries have been examined. ORG estimates that Home Office experts checked only 0.3 percent between April and July, based on the department's projected monthly caseload of 17,515. ORG also questioned whether the tools were saving much time. It says the average monthly number of initial substantive asylum decisions since April 2026 fell about 42 percent compared with the same period in 2025, while the mean time from interview to decision rose 29 percent. While those figures don't prove that AI caused the slowdown, they offer little evidence of the productivity gains ministers have promised. Oversight of the separate Asylum Policy Search tool also appears lacking. The Home Office confirmed in August that expert testing had stopped, despite previously telling MPs that checks were continuing. ORG says no summary of the testing results was produced. Meanwhile, asylum applicants are not specifically told when AI has been used in their cases and cannot inspect or correct the summaries generated from their interviews. ORG also questions the Home Office's claims of human oversight. The department's own data protection assessment says ACS automates work previously done by technical specialists, who reviewed and summarized asylum cases before passing them to decision-makers. ORG warns the tool could remove an experienced layer of scrutiny, rather than simply save staff time. The group is calling for both tools to be suspended and independently audited, arguing that the current safeguards are inadequate for decisions that determine whether someone can remain in the UK. It's not the first time the Home Office's technology plans have come under fire. In June, more than 60 rights groups urged ministers to abandon plans for AI-powered age estimation of asylum-seeking children, warning that the technology was unreliable around the crucial 16-to-18-year-old age range. Later that month, an inspection found that the department's long-delayed Atlas immigration system was failing to provide caseworkers with feedback from asylum appeals, forcing staff to rely on workarounds. The Home Office did not respond to The Register's questions. ®

VMware has discovered so many flaws in its products using AI that it felt the need to create a new product to help users manage them. Prashanth Shenoy, vice president of product marketing in the VMware Cloud Foundation (VCF) division of Broadcom, told The Register that VMware recently moved to deliver monthly software updates, in part to address bugs found by AI. He admitted, however, that users struggled to keep up with the pace of change before the move to monthly releases because it takes time to test the impact of patches – even though VCF allows non-disruptive live patching. Shenoy said customers therefore need "room to breathe." Enter a new tool called HyperArmor, which does three things. One is "virtual patching" – blocking attempts to exploit a vulnerability without changing the vulnerable software itself. VMware already offers virtual patching in its vDefend network security product and explains how it works by referring to CVE-2024-5806, the bug in file transfer tool MOVEIt that made waves in 2024. VMware says vDefend can detect patterns associated with attempts to exploit the authentication bypass flaw and block the traffic. HyperArmor will do similar things to block malicious traffic heading toward VCF implementations. Another job HyperArmor takes on is rejecting any actions that alter known good states in VCF. If a human approves a change, the tool will update policies to allow the altered state. The third job is multi-factor authentication (MFA). The Register put it to Shenoy that MFA is now table stakes for security. He agreed but explained that VMware designed HyperArmor to protect against attackers using stolen privileged credentials to access VMware's ESX hypervisor. To (hopefully) stop attackers using those creds to take over a host, HyperArmor requires users to input SSH credentials and a secondary time-based one-time password sent to a user with a verified identity. HyperArmor is a discrete product that VMware will sell as an add-on to VCF. That means VMware thinks its customers are going to be OK paying extra for MFA. While VMware will sell HyperArmor, it didn't create the product. Broadcom's security businesses Symantec and Carbon Black cooked it up. Broadcom's CA business (formerly Computer Associates) has done similar things. At VMware Explore on August 31, Broadcom launched AgentMinder, which aims to stop AI running inside VCF going rogue. AgentMinder came from CA but was launched as a tool for VCF (and third-party Kubernetes stacks). At VMware Explore, The Register met Ram Velaga, president of Broadcom's Infrastructure Software Group, and asked him if the group has a single strategy, or if its divisions each do their own thing. Velaga said Broadcom uses a combination of both approaches. "There's a level of focus with individual teams that you don't have if you have one very large central team," he said. "You get the speed at which you have to innovate with individual teams." Broadcom's Infrastructure Software Group therefore operates with at least six teams, and when one of those teams wants to add functions it will look for existing assets developed by another internal team rather than trying to write its own code. For broad products like VCF, the work of several Broadcom teams could therefore be relevant. The Register has no idea which team created the name HyperArmor, a question we pose as Cisco already offers a "Hypershield." Maybe this sort of name is what's needed to stop a HyperHoodie? ®

Microsoft took down the Exchange Admin Center (EAC) today after a code change left administrators facing an "Internal Server Error" when trying to access the service. Microsoft first admitted that something was amiss at 0937 UTC on October 12, noting, "When attempting to access the Exchange admin center, affected admins may encounter an 'Internal Server Error' message." Any administrator attempting to access the EAC could be affected. It was more than an hour before the company stated: "We've identified unexpected 302 redirect errors during affected Exchange admin center (EAC) access attempts and are analyzing diagnostic logs to isolate the source of the access failures. "While we continue our investigation, admins remain unable to access the EAC, with no workaround currently available." So, a fun morning for Microsoft's European users. According to Microsoft, the EAC is "a modern, web-based management console for managing Exchange Online." Want to manage mailboxes? EAC is where you go. Group management? Again, EAC. Microsoft explained: "The design of the EAC in Exchange Online aligns with the overall Microsoft 365 admin experience." Sadly, that alignment appears to extend to service wobbles caused by changes rolled out without due care and attention. Two hours after admitting things had gone awry, Microsoft said it was "continuing to test the rollback of the recent Exchange admin center (EAC) change while monitoring the results to determine if this resolves the impact." At this point, administrators were nodding sagely at an all-too familiar story. Of course rolling back the change did the trick. Microsoft stated as much in its final update: "A recent EAC configuration change resulted in unexpected HTTP responses within the affected request path, preventing admins from successfully accessing the EAC. "Reverting the identified change restored expected service functionality." Microsoft also assured customers it was reviewing the change to ensure it didn't make the same mistake again, as well as its monitoring and telemetry, to "improve detection and alerting for similar service behavior." The problem is that this is hardly a new phenomenon. In September, Microsoft made a configuration change that left some users unable to access SharePoint Online sites and pages. In April, another ill-thought-through change took down Outlook for iOS. And, in 2025, Microsoft 365 tumbled from the cloud because of yet another configuration change. It's almost as if testing changes consists of little more than asking Copilot, "does this look alright to you?" and then hitting the go button. The latest outage was resolved before the end of the working day, and a preliminary Post-Incident Report is due within two business days, with a final report due next week. Hopefully one or the other will contain words along the lines of "Hey, this keeps happening, maybe we should do some more testing before deploying to production?" ®

If you thought everyone was into chatbots, think again: about half of US adults never use the AI tools, and most have no interest in starting. According to Pew Research Center, older adults are the least likely to use chatbots, with roughly three-quarters of those aged 65 and older saying they never use them. That may be expected, yet more than half (58 percent) of adults aged 50 to 64 don't use them either. Those aged 18 to 29 use chatbots the most, at 66 percent, however even a third of them say they never do. About four in ten under 50 report not using them at all. Across all adults, 51 percent are chatbot refuseniks, so users are - just barely - the minority. Pew asked non-users to choose from five possible reasons. Six in ten cited simple lack of interest as a major factor, followed by the more familiar concerns of data privacy and the accuracy of chatbot answers. Some Reg readers might suspect older people avoid chatbots because they lack tech-savviness, but fewer than a third cited not knowing how to use them as a major reason. The authors of the survey also found fewer than one in five worried about what others might think of them for using the technology, so social stigma isn’t the deterrent either. Adoption also differs by race and ethnicity: non-users are more likely to be white, black or hispanic than Asian, according to Pew’s survey. The AI dodgers don't look ready to budge. Two-thirds (67 percent) say they’re unlikely to use the technology in the next 12 months, including 40 percent who say they’re not at all likely to change their mind. Another 12 percent are unsure, and just 5 percent of non-users think they'll probably try a chatbot over the next year. Perhaps abstainers were put off by a bad experience, such as Microsoft’s notorious racist chatbot Tay a decade ago. Or something more recent, like the Trump administration’s official America.gov service, which at the time of writing still displays a hotdog if you type that into the prompt. ®

What is DNS? That's right, it's a database. You can disagree with me but you would be wrong. What can you put in a database? Strings, ints (and you learn the difference right quick if your AWS account ID has a leading zero!), JSON objects, coordinates; some of the databases out there have some pretty esoteric types. PostgreSQL supports circle, a polygon, a macaddr8, and a tsquery. Oracle will store a ROWID and your dignity. In a word, databases store data. Can you put answers in a database? Well sure; you could record that "3 + 2 is equal to 5" and store that answer. You can do the same with comparison operators; "3 is larger than 2" and so on. Take it a step further and you've built something with your database: a remarkably inconvenient calculator. So now we've established that you can store arithmetic and decisions in a database. And what can we do with those things? That's right, we can follow instructions! Good job! And what does following instructions require? Generally, memory; a place to store your numbers, your instructions, and which instruction comes next. If only we had a place to store those things. Oh right, we do; we have a database! Now connect those lookups together. Read an instruction, look up an answer, save the result, repeat as needed. So what have we built? That's right: a computer! And what can a computer run? Programs! My, you're quite good at this. And now, gentle reader: what program do we apparently have a moral obligation to run on every computer? Oh yes, though many of you are going to sit back and instead say "oh no" as the horror courses through you. I am talking about Doom. This very publication agrees with me! What have I done? That's right; I have compiled Doom for a 32-bit machine whose arithmetic and memory live in DNS. A small program supplies the wiring and the clock, neither of which the architects of DNS saw fit to include. But DNS does supply the answers. So, can we play Doom? Only if it's turn-based. It won't render on human timescales anyway. In my first attempt, I pressed Right, and then waited for 577 seconds to find out what happened. By using native transport, pipelining, packing registers densely, and optimizing instruction blocks, I was able to drop that to 146 seconds, before I hit a wall. BIND answers in about 7 µs per query over TCP and 4 µs over UDP, which is fast. The problem is that a 32-bit add is four DNS lookups, a branch is five, and a frame needs 6.6 million of them, so don't use a server that charges per request. And what do we do when BIND isn't performing the way we want it to? I don't know what you do, but I publish my code, then reach out to Paul Vixie, DNS pioneer and architect of BIND 8 to unjustifiably complain. Paul then apparently congratulates me on "the best use of RFC 2136 DNS UPDATE" that he'd ever seen. Given that he's a co-author of that RFC, I suspect he's seen a lot! Which was gratifying, but unhelpful. So let's take another crack at it. What if I do lookaheads? Good thought, bad outcome, because BIND is bound by message count, not chain depth. Well okay, I have some more tricks in my Big Bag of Stupid Ideas. So I went a bit further and via committing registers to DNS less often, moving lookups to UDP, telling BIND to stop padding answers with authority records, and requiring fewer lookups per instruction, I was ultimately able to get that press-right time down from 577 seconds to 146.3 seconds to 54.4 seconds, for a 10.6x improvement. And now I think I have hit a wall. Even at one DNS exchange per instruction (which is far better than anything realistically achievable), 1.4 million instructions at roughly 10 µs each is 14 seconds per tick. Doom says its tic rate is 35 per second, so "playable" is a tic every 28.6 ms. I remain at 54,000 ms, which is three orders of magnitude away from victory. Okay seriously, WHAT? This legitimately works. Every commit is an RFC 2136 UPDATE with a hard requirement: it fails unless the previous state is still exactly what was last written. Apparently "real" database people call that optimistic concurrency, or compare-and-swap. Joke's on you, friends; DNS has shipped with that since 1997; has your inferior database? The database itself is 6.4 million rows of truth tables. ff.01.0.add holds "001": 255 plus 1 is 00 carry 1. DNS is just acting as a librarian here, fetching memorized answers to any arithmetic question in-scope and returning the answer upon request. It's also verifiable! The frame that's spat out is checked on every pixel against an independent emulator. If one record is corrupted in the AND table, the answer changes, which proves there's no local fallback doing arithmetic for us. I've included tooling so that reviewers can run two commands against that repo and verify correctness. I also confess I cheated slightly here; the run I'm pulling these numbers from started from a checkpoint prepared by a normal emulator. It turns out that for SOME UNKNOWN REASON getting a cold boot through Doom's startup screens via DNS would take hours. Also, before the timer starts, DNS has to load those 6.4 million records, so it's a computer that takes roughly a minute to boot before it can even start. We can't cheat This runs atop nothing special: Debian's stock BIND 9.20, configured through named.conf. It lives in a container with networking disabled. I want to be very clear here: this isn't DNS pretending to be a CPU. The instruction decoding is done ahead of time by a compiler; BIND never executes an instruction itself. All that it sees is a key-value read, a conditional write, and 6.4 million rows of precomputed answers. So instead, think of this monstrosity as a database with a key-value read, a conditional write, and 6.4 million rows of precomputed answers. It's not really a computer; it's simply a very good database. And that turns out to be enough. ®

AMD is expanding its rack-scale AI lineup with a design to run AI agents, offering CPU-only configurations or integrated CPU/GPU deployments, and liquid-cooled or air-cooled versions. Revealed at the Open Compute Project (OCP) Summit 2026 in San Jose, California, the chipmaker's new rack reference design is named Helion, while its existing system carries the moniker Helios, which we can't help feeling is inviting confusion a little. Helion offers an OCP-compliant blueprint for agentic AI infrastructure built around AMD's CPUs and accelerators. AMD launched Helios earlier this year for frontier-scale AI training and large-scale inference. Helion's enterprise configuration instead combines agent execution and inference within a single rack. Customers can opt for a CPU-only hyperscale configuration or an enterprise configuration with a mix of CPU and GPU blades. Depending on configuration, AMD advertises up to 32,768 of AMD's "Zen 6" CPU cores, delivered via compute nodes fitted with 256-core versions of its Venice Epyc processor, up to 16 of its Instinct MI350P GPU accelerators, and up to 102 Tbps of networking bandwidth. The CPU-only hyperscale version offers dedicated capacity for agent execution and connections to separate model inference infrastructure, while the enterprise configuration combines agent execution, GPU inference, and supporting workloads in one rack. Both configurations offer dense, balanced, or high-performance CPU blades. For dense deployments, EPYC 9006 chips rated at 400 W are used, while performance blades sport a version of the chip with just 96 cores but clocked at 5 GHz. Balanced blades use 256-core processors rated at 600 W. There are also GPU inference blades, with eight Instinct MI350P GPUs in a 5U server, plus general-purpose blades, intended for database and storage servers or technical compute workloads. AMD SVP for Compute and Enterprise AI Dan McNamara said agentic AI is not a single workload, spanning enterprise applications and complex workflows, each with different compute requirements. "To do that takes optimized compute and infrastructure that can adapt to different workloads. With AMD Helion, we are bringing Epyc performance, open software and flexible rack designs together so our partners can build systems around customer needs and make efficient use of available power," he added. AMD says it plans to contribute the Helion specifications to OCP, making them available to the community for further development and adoption. Partner-built Helion configurations are scheduled for November 2026. In its announcement, AMD cites Cisco, Dell, HPE, Foxconn, Lenovo, Supermicro, and others as on board with the new kit. ®

EVENT: Every sovereign AI architecture is a trade, and the currency changes with the option: capital, capability, elasticity, or the fine print about who operates it. The cost you notice last is usually the one that decides whether the pilot ever becomes production. Nobody chooses badly on purpose. They choose with the information available at the planning stage, and the costs that matter most, in capability and in operations, tend to show up later. Which is why the honest comparison is hard to get. The people who have run each option for a year are the ones who know, and they are not saying it on a recorded webinar. They will say it at dinner. On Thursday 29 October, just over two weeks from now, The Register is hosting HPE and NVIDIA in Washington with editor-in-chief Matt Rosoff in the chair and the Chatham House Rule in force. HPE sends Bharath Ramesh, Head of Product, AI Factory Solutions, and Tracy Mills, Acting Director Federal Systems Integrators & Distinguished Chief Technologist Federal. NVIDIA sends Bert Condensa, Vice President for Enterprise AI Factory Segment Sales. All three are there to take questions. The agenda is the trade-offs themselves: which infrastructure options hold a boundary, what each one costs in capital, capability or elasticity, and whether the answer survives the move from planning into operations. And the questions underneath: what can leave a jurisdiction versus what can leave the building, and what an auditor will accept as evidence a year from now. This is written for the person who owns that decision and is making it this quarter, not next year: technology, data and risk leaders in healthcare, financial services, research, pharmaceuticals, government and higher education. It is one table with no slides, everyone at it a peer, and a chair who closes down a pitch as soon as it starts. Places are filling. Dinner runs from 6.00pm to 8.30pm at Fiola Mare, Washington DC. Request one here. Government and public sector guests: where your agency’s ethics rules require it, you can pay your own share and we will provide a receipt. Please say so when you request an invitation.

Microsoft has closed a loophole that let users sidestep the need for an account during setup for consumer versions of Windows 11. If users clicked "Learn More" when the setup process prompted for a Microsoft account, the Out-of-Box Experience (OOBE) skipped directly to the point where they just had to enter a name and password. Windows prodder "Bob Pony," who first documented the loophole, reported its removal. We tried the original trick on real Intel hardware and found it was possible to skip the Microsoft account requirement by clicking a link, but no longer. Other hacks – for those who dare to use unofficial Windows setup routes – still exist, though they are beyond the skill set of the average user. Microsoft requires an account during initial setup of Windows 11 Home and Pro for personal use. For consumers, using Windows 11 without an account can be tricky because services such as OneDrive require authentication. Microsoft is far from the only operating system vendor to demand an account during setup, but many Linux distributions, by contrast, don't require one. According to Bob Pony, the loophole was "quietly patched server side," which makes sense. Microsoft was hardly likely to leave it flapping open for long. Microsoft has won this round of account workaround Whac-a-Mole, though another bypass may yet pop up. ®

Yandex engineers are working to restore services after Ukraine launched drone strikes on three of its five Russian datacenters. The strikes began overnight on October 7-8, "seriously damaging" Yandex's Sasovo datacenter and leading to a total shutdown. Ukrainian drones struck again on the morning of October 9, targeting the company's Kaluga datacenter, resulting in part of its infrastructure being taken offline. Dima Pirogov, head of Yandex's fintech division, confirmed via Telegram on Sunday morning that a third strike targeted the company's Vladimir site, which supports Yandex Pay. A day later, services remain offline, although Pirogov said users' money is safe. "I'm very sorry I can't tell you yet that everything is working again," the exec wrote. "The team continues to restore services and is checking every step. When it comes to your money, it's important to us not just to restore Yandex Pay as quickly as possible, but also to ensure that everything is working reliably. "I understand that you're not waiting for words, but for access to your money. And I understand how hard it is to wait when you can't use it. We can't give you a precise timeline for the restoration yet, but we promise: we won't stop until the service is back online." Yandex Pay is the company's digital payments service. Users can make online and in-store purchases, pay in installments, and split bills with friends and family. It's also integrated into online shopping experiences. It is comparable to Apple Pay or Google Pay, while Yandex more broadly offers services similar to Google's. Yandex said engineers are prioritizing the most-used services, and reported that access to the Alice AI assistant, Mail, Maps, Search, and Taxi is being restored, but other features remain offline. The company's status page lists widespread outages across Yandex Cloud, its public cloud platform, following the attacks. Updates from the maintenance team describe partial availability despite the outage indicators. They suggest that some customers may be able to access certain features, such as Yandex Cloud Console and API, but availability remains inconsistent across the platform. In its most recent update, the company said: "The Yandex Cloud platform is still operating in a limited mode. Our specialists continue restoration work around the clock. The technical support team receives an increased number of requests, so the response time can be increased. "Our recommendations continue to be relevant. We recommend using alternative sites for hosting and reserving resources." Internet monitoring organization Netblocks reported that Yandex Cloud's network connectivity was at 30 percent of normal levels on October 11. Ukraine president Volodymyr Zelensky said the drone attacks were a response to Russia's own, which had targeted datacenters in Kyiv. The BBC reported that in September, Russian drone strikes targeted a business center in the capital city, killing four people and damaging a datacenter, which in turn resulted in around 100,000 households losing internet access. The attacks led to other knock-on effects, such as public transport boards not displaying information in Lutsk. Kyiv's attack on the Vladimir site on Sunday morning came hours after controversial comments from US president Donald Trump, who accused Zelensky of wanting "to make problems for the world" by continuing to attack Russian oil refineries. Trump went on to say that it was time for Ukraine to elect a new president after Zelensky said measures should not be taken that allow Russia to generate more revenue to fund its war. Zelensky was referring to a deal between Trump and Putin to release diesel from Russia to the US and the rest of the world. Later that day, the US president claimed Russia and Ukraine had agreed to an "energy ceasefire" to halt attacks on each other's energy infrastructure. Kyiv said it was unaware of the agreement before Trump announced it on his social media platform, but was willing to halt attacks on Russian energy infrastructure if Moscow reciprocated. Russia launched a missile attack on a Ukrainian power plant hours later. ®