The Register

You'd swear from the over-the-top reactions that someone is threatening to take European Commission workers' dogs out back and shoot them. Why the freakout? Because the EC is deploying a self-hosted alternative to Microsoft Teams to keep staff communicating if Microsoft's service becomes unavailable. To be clear, the Commission isn't replacing Teams with Element Pro; it's setting up a backup. That's it. That's all. But according to a Politico report, users are having fits. Element is "absolute shit," rants one. Another proclaims: "The US would 'instantly win the war' if it ever decided to switch off its technology." Uh, guy, that's exactly the point. If US providers cut off European users, whether by choice or government order, many EU governments and companies would face serious disruption. That's exactly why Europe needs an open source, sovereign backup. That's the entire point of this exercise. Take away all the hysteria, and you're left with a rather mundane business-continuity initiative. Don't think for a minute that Trump wouldn't have a snit and order tech giants to block services to officials. Just ask Kimberly Prost, a Canadian International Criminal Court judge, who, besides losing tech services, also lost her credit cards, access to Amazon and Google services. "Alexa wouldn't talk to me," she says – punishment indeed. Now imagine Russia and its allies start attacking the Baltic states. Oh, wait, Putin is already doing that. Then it gets heated, and Trump decides to side with his good friend Putin. What happens to American tech companies' services in Estonia, Latvia, and Lithuania? I can answer that in two words: "Nothing good." That said, the EC of course is not dumping Microsoft services. No, it has configured Element Pro to run on the Commission's own infrastructure. An internal memo described the service as an "internal backup" that gives the EU's executive branch an alternative channel for internal messaging, audio and video calls, file transfers, and project discussion rooms. Let me spell it out for you. This is a Commission initiative. It's not an EU-wide order to abandon Microsoft software. I'd love that myself, but that's not what's happening here. Teams will remain the principal platform, while Element provides a fallback intended to reduce dependence on an externally operated communications service. Got it? I hope so! Mind you, there are other good reasons to dump Microsoft Teams. I'm not a fan. For my sins I have to use it and every other groupware program out there. And, I might add, I've been using them since Lotus Notes rolled out the door in 1990. Digital sovereignty isn't just a good idea. It's essential. Or, as an EC July staff memo put it: "In practice, this means having greater control over our data and systems so that we can reduce risks and limit disruption to our work, wherever possible." Exactly so. Element's architecture is well suited to digital sovereignty. Built on Matrix, an open standard for real-time communications, the platform supports self-hosting, decentralized operation, and interoperability with other Matrix-based systems. Element also supports end-to-end encrypted messaging and calls. You see, the sovereignty argument isn't just about getting an EU-based technology stack. After all, Element is a British company. No, what really makes Element a winner for the EC is that it's mature groupware that lets them deploy and run the system on their own infrastructure rather than relying on Microsoft's cloud. As for the crybaby users, suck it up. No one ever likes change. If the EC switches over to Element lock, stock, and barrel, I suspect much of today's outrage will be forgotten in six months. I've seen this time and time again. Some short-term inconvenience is a reasonable price for greater control over essential communications in an uncertain world. A complete switch isn't the plan now. I think it should be. ® Bootnote Element told The Register: "Governments and defense organizations are migrating to sovereign alternatives to ensure they can still operate effectively in extreme circumstances. As IT strategy and IT functions move quickly, less adaptive end-users may well feel discomfort as the US Big Tech solutions they've been used to are replaced by sovereign alternatives. Hopefully this is the largest inconvenience most people experience over the coming decade. "We've only had positive and constructive feedback from our actual customers; the leaders and IT functions that are driving sovereignty. "There's inevitably a few unhappy change-resistant users in large deployments, but we work in lockstep with our customers to gather feedback and incorporate it, especially during the trial phase. From what we see in the field, open source alternatives like Element are ready for primetime as a much needed alternative to dependency on the big tech vendors."

Nearly a third of developers who use AI daily spend four hours or more of their working day with it, according to Stack Overflow's latest survey. The survey received 30,903 responses from developers and technologists across 169 countries. Among the 10,926 daily AI users who answered the time-use question, about 80 percent reported using the tools for at least an hour a day. The research by developer Q&A site Stack Overflow also found skepticism from developers toward using AI in their roles. Only about 7 percent of developers said they trusted AI output for important work decisions, while 48 percent trusted it when they could easily verify the output. Seventy-nine percent considered source attribution important or very important when deciding whether to trust an AI-generated technical answer. Stack Overflow CEO Prashanth Chandrasekar said developers were highly skeptical as a group. "As AI agents become rapidly more autonomous, that skepticism now acts as a critical safeguard. As agents take on more work, developers need to know where an answer came from, what context informed the information that was produced or a decision that was made, and the most crucial piece – that they can actually verify its accuracy and relevance. Trust is not a given when it comes to developers; it is earned through transparency, source attribution, and rich technical context." The survey found that coding assistants or coding agents are used by 66 percent of respondents, general-purpose chat tools by 63 percent, and automated agent workflows by 26 percent. Seventeen percent of respondents said they did not use AI tools. Stack Overflow's commercial relationships with AI companies have proved contentious. In early 2024, it introduced OverflowAPI, offering paid access to its question-and-answer content and requiring participating AI providers to attribute answers to relevant posts. A couple of months later, users began to revolt against its partnership with OpenAI, which allowed the LLM builder to train AI models using 15 years of users' contributions. The Q&A site was also set to benefit from the deal by getting access to OpenAI's technology. A number of users said they'd rather remove their posts and risk their reputation scores than have their contributions train OpenAI models. The survey also found that 52 percent of respondents had learned a new coding skill or programming language this year. AI code-generation tools were among the most common learning resources, behind technical documentation. ®

Microsoft has joined the Jev fan club, an accidental group of companies that share a common desire to be recognized for their own decision models. Jev, announced by TypeSafe AI three weeks ago, is a large language model (LLM) tuned to respond to certain types of questions with a limited range of responses, rated by probability. Due to its speed, relative affordability, and response constraints, it's well-suited for a variety of business applications where open-ended text of uncertain accuracy might be undesirable. One of the selling points of Jev is that decision models don't hallucinate in the way that standard LLMs do. But decision models can make errors and their popularity has already prompted researchers to explore how those errors might be magnified. Nonetheless, they have their uses. The attention lavished on Jev prompted other companies to declare that they too have decision models to offer, even though machine learning researchers have long been able to create classifier models for probability-based decisions. OpenAI said its Decisions API has entered public beta. Cloudflare chimed in with its Clef model. Strands trotted out Strands Decider 2B, "a small, open source, decision model." Liquid AI introduced d1. Perplexity launched its Decisions API. Snowflake talked up its own decision model. Then there's Surogate Rune and H2O.ai's H2O-Lightning-4B, a decision model built on Qwen3.5-4B. All told, more than 100 such models are now vying for attention. Now it's Microsoft's turn. "Decision models are quickly emerging as an important new category in AI," said Achint Srivastava, VP of software engineering in the Office of the CTO at Microsoft, in a blog post on Friday. "Unlike LLMs, which are designed to generate text or reason through complex problems, decision models are purpose-built to deliver structured outputs that software can immediately act on. And once you understand that capability – making decisions and classifying things at very low cost with high performance – all kinds of useful tasks get unlocked." Microsoft's entrant into the race is called Microsoft-Decision-1, which is offered via Microsoft Foundry and, soon, via OpenRouter. Redmond's decision model, like H2O's, is based on a Qwen model, Qwen3.5-9B in this instance. The Qwen model family is developed by Alibaba Cloud, the cloud computing arm of Chinese tech giant Alibaba. Microsoft, for reasons not disclosed, said it will soon rebase the Decision-1 on its models and those from OpenAI. The US software giant claims its model is 2.5x faster than H2O-Lightning-4B and 2.8x faster than Jev in its latency test, leads the pack in accuracy (83.5 percent) on 36 benchmarks, ranks second (behind Quyet-1.0-Large) in confidence score (92.2 percent), and is more than 20x cheaper than OpenAI's GPT-6 Sol in text classification tasks. Input tokens cost $0.042 per million tokens and output tokens are free. "Now that agentic AI is a reality, we’ve seen that cost plays a major role in how people decide to use AI," said Srivastava. "And it’s increasingly important to choose the right model for the right job." ®

Bob, possibly the same Bob whose conversations with Alice draw so much interest from eavesdropping Eve, was browsing a site we'll call TechHub. The site hosts an AI agent served by Amazon Bedrock AgentCore. Bob asked the agent for help understanding the content of a URL, a credential endpoint – in raw JSON, if you don't mind. The endpoint returned data from the Instance Metadata Service (IMDS), which provides metadata about cloud instances and VMs at providers like AWS, Azure, and Google Cloud Platform. The metadata includes details like region and availability zone, subnets, system images, security groups, public keys injected during spawning, but also potentially more sensitive details like user data and security tokens. IMDSv2 addresses some of these risks, but back when Bob was browsing late last year, Bedrock AgentCore still used IMDSv1. The metadata provided to Bob by the helpful agent contained the agent's temporary credentials. So Bob loaded them onto his local machine, and remotely enumerated the company's other agents in that AWS region. He then logged into the Amazon Elastic Container Registry (ECR), pulled the agent container images, and ran each as root to inspect the source code. The stolen credentials also allowed Bob to discover the memory resources available in that AWS region, including the ones used by agents. From these, he's able to extract the users and their agent sessions – their conversations. Researchers at Zenity Labs disclosed their findings to AWS in December 2025. "We discovered that agents deployed through AgentCore could access their instance's IMDS endpoints," said Tamir Ishay Sharbat and Lana Salameh in a blog post. "This meant that an external attacker with nothing more than chat access to a single exposed agent could send a single prompt, extract its IMDS credentials, and use them to take over all AgentCore agents in the same AWS account and region." The basic problem, they explain, is that the Firecracker MicroVM used by AgentCore failed to provide sufficient network isolation. So an attacker – I know you're thinking it was Mallory, but it was really Bob – could direct the agent to carry out a server-side request forgery (SSRF) attack by fetching temporary AWS credentials for the IAM role assigned to the workload. And because the default AgentCore role was overpermissioned – it was scoped to all AgentCore resources in the region rather than a single agent – anyone in possession of the temporary IAM credentials could launch other agents, read sessions, write agent memories, and fetch secrets from AWS Secrets Manager. "By leveraging the IMDS credentials we could send direct API requests to create new memories across different agents and users," said Sharbat and Salameh. "These in turn would persistently alter agent behaviour and hijack the agents’ goals across future sessions." The Zenity researchers told Bob's tale, or something like it, to AWS last December, then followed up in January 2026 with details about AgentCore being overprivileged. On April 12, 2026, AWS responded to the security biz that its report was "informative" and closed the report, noting that as of February 14, 2026, AgentCore had been updated to use IMDSv2 exclusively. AgentCore's excessive permissions remained in place at least until June 22, 2026, when Zenity checked in and found the issue hadn't been remediated. A final review by Zenity occurred on September 29, 2026, at which point the biz observed that AWS had addressed the extant problems, clearing the way for Bob's hypothetical adventure to finally be told. ®

Microsoft 365 subscribers on Family, Premium, or Pro tiers who make full use of their OneDrive storage allotments consider this a warning: It’s fixin’ to shrink. Redmond published a support page this week detailing changes to Microsoft 365 shared storage allotments, and while the company is painting them as offering greater flexibility for users, they also represent a massive reduction in capacity. The former storage allotment schema offered 1TB per user on all three of those tiers, with a cap of six users, for a total of 6TB. The new storage quotas will be shared among all users on those accounts, still with a six-user cap. Family or Premium plans now get 2TB of storage space to share among up to six users. Pro tier users will have to share 5TB. In effect, that means fully shared Family and Premium subscriptions will lose up to 4 TB of combined storage capacity, while Pro subscriptions will lose up to 1 TB. Some existing customers may qualify for temporary bonus storage. On the plus side, there's no longer a fixed 1 TB cap per person, provided the shared pool has room. “The goal is to provide a more flexible experience,” Microsoft offers as its justification for the move. “Shared storage helps ensure that one person doesn't run out of space while other members have unused storage available.” That might be great for heavy users on an account, who can now go over a 1TB cap, but if, say, four people on a family plan are all using 500GB and someone wants to upload something else, that’s too bad. You can always buy more OneDrive space, of course. Microsoft appears to understand that users might see this as a reduction in storage, posing an FAQ question: "Is this change a reduction in storage?" The answer again highlights "flexibility," and helpfully notes, "For example, if the subscription owner doesn't share the plan, they can use the full shared storage allotment." (Individual plans already exist but top out at 1TB.) Microsoft also appears to understand that the move might cause a backlash, and has included subscription cancellation steps right on the support page for the change. New Family, Premium, and Pro tier subscribers as of Thursday will see this new storage allotment schema. Existing subscribers will be hit with the change when their plans renew on or after May 2, 2027. Microsoft said it will remind users of the change as their renewal date approaches. Customers are not thrilled about this move. A thread on Reddit discussing the change has nothing positive to say about it, with commenters noting that they’re likely to cancel their subscriptions rather than let them renew and face a storage reduction and a need to clean up files. Much of the Microsoft365 subreddit, as of writing, is filled with posts complaining about the move, with one of the top posts reminding people to turn off their automatic renewal to avoid getting hit with the storage slash. The Windows maker has a semi-regular habit of slashing cloud storage allotments for its 365 customers. Office 365 used to come with unlimited cloud storage, but that offer ended after some users actually took unlimited to mean unlimited. Storage tiers have been further reduced over the years, most recently in 2024 when McGill University drastically slashed students' OneDrive allowances to 20 GB following Microsoft's changes to education storage allocations. In other words, don’t get too comfortable with 2TB of shared storage for six people – that deal, too, may be altered further. ®

SpaceX is looking to take on the incumbent mobile networks in the US with a combination of low-band spectrum that can penetrate indoors together with the 2 GHz spectrum supported by its Starlink Gen2 satellites for bandwidth. The rocket-flinging operation says it cut a deal to acquire a nationwide low-band spectrum license portfolio of up to 14 megahertz of paired spectrum in the 800 MHz band, which will pave the way for Starlink Mobile to become a major cellular carrier in the US. SpaceX agreed to acquire the spectrum licences from telecoms investment firm Grain Management, which bought them from T-Mobile. The transaction is subject to FCC approval and other customary closing conditions. In its FCC filing [PDF] regarding the transaction, SpaceX says it intends to roll out its own terrestrial infrastructure for this purpose, operating standard wireless equipment such as antennas on towers, rooftops, and other structures to provide reliable signal coverage. Financial terms of the agreement were not disclosed. We asked SpaceX for further details and will update if we get answers. Also this week, America’s telecoms regulator, the Federal Communications Commission (FCC), approved Starlink’s application [PDF], filed last year, to put into orbit an additional 15,000 satellites that are optimized for 2 GHz spectrum globally. The operator says this will allow it to bring high-speed service directly to unmodified devices anywhere in the world. It builds on SpaceX’s $17 billion purchase of spectrum from satellite biz EchoStar last year, which netted it 50 MHz of exclusive S-band spectrum in the US, as well as the firm's global Mobile Satellite Service (MSS) licenses. Taken together, these two moves could pave the way for Starlink Mobile to build a hybrid satellite-terrestrial network, with the low-band spectrum helping to reach subscribers inside buildings, while the 2 GHz spectrum used by its satellite constellation could extend high-speed coverage to areas beyond the reach of cell towers. It is understood that most existing mobile devices already support the 800 MHz band frequencies, which SpaceX describes as currently underused, meaning subscribers can connect without having to upgrade. “Starlink Mobile can now become the first network operator to deploy both satellite and terrestrial spectrum. One network that provides Americans with a reliable service indoors, outdoors, in cellular dead zones, and everywhere in between,” the firm stated in an update. According to finance news site MarketWatch, shares of the incumbent wireless operators fell following the news. It says that Verizon shares were down almost 6 percent, while AT&T and T-Mobile US shares were down more than 6 percent. T-Mobile kicked off a Direct-to-Cell service for US customers using Starlink’s existing satellite constellation last year, offering it even to customers of rival networks for $15 per month. A report out earlier this year found that the number of direct-to-device connections rose nearly 25 percent between July 2025 and March 2026, but a separate report cast doubt on how widespread adoption of the technology will be. Juniper Research pointed out that it is inherently niche, kicking in only when the terrestrial signal fails, and it does nothing to solve poor indoor coverage – something that SpaceX’s 800 MHz move aims to address. ®

Large-scale AI training has already escaped the confines of a single datacenter. Google said Gemini was trained synchronously across clusters in multiple locations; Microsoft has connected AI data centres in Wisconsin and Georgia into what it describes as one distributed AI supercomputer; AWS has connected AI compute clusters across wide areas to allow Anthropic to build Claude models; Meta has built high-capacity datacenter interconnects to support model training; and CoreWeave and Google Cloud recently announced cross-cloud training with a private interconnect, with Azure likely to follow later in the year. This growing geographic spread of model training is partly due to the limits of single datacenters or campus clusters, which can be strained as they try to meet the compute and power demands of new models. Cisco estimates that training models today can require clusters with tens of thousands of GPUs. By 2030, the largest individual frontier training runs could draw 4-16GW of power, according to researchers at Epoch AI. Distributing compute allows hyperscalers and datacenter operators to build in locations with more available power or space and fewer planning constraints. How to train an LLM At the core of an LLM is a neural network with billions of numerical parameters, or weights, that are adjusted during training. A batch of training data passes through the model, which makes a prediction. The system measures the error and calculates how the weights should change. The weights are updated and the process starts again. At the scale of modern models, this work is distributed across thousands of GPUs and other accelerators, such as AWS Trainium and Google TPUs. Different GPUs can process separate batches of data or different parts of the model. But they cannot work entirely independently: at various points they must exchange results and synchronize before the next stage of training can begin. What networks carry between the accelerators is generally not the text or images used to train the model, but large arrays of numerical data such as gradients and intermediate results. Thousands of accelerators may need to exchange and combine that data at roughly the same time. "If you look at any training job, it is basically a repetition of compute and synchronization," explains Ramesh Sivakolundu of Cisco's Silicon One architecture team. "You compute the weights, communicate those weights to synchronize every GPU in the cluster, and then restart the computation. That means bursty traffic happens periodically throughout the training process." Thousands of accelerators can finish a computation phase and start communicating at the same time. This can lead to "incast" problems where many senders converge on the same destination and traffic arrives faster than the next link can handle. The fundamental issue lies with synchronization. In a synchronous training job, one group of GPUs cannot simply ignore a slower group and continue independently. The network-wide exchange must complete before the participating GPUs can move on to the next stage. A network delay therefore becomes compute delay. "You don't want the network to be the bottleneck in the training process," says Sivakolundu. "The idea is to keep the GPUs fully occupied and not have the network introduce additional latency into the overall training program." That does not mean that every instance of network congestion or packet loss brings the entire model training to a complete halt. However, a delayed or dropped flow may require retransmission, prolong a collective communication operation, and leave other accelerators waiting. More serious failures can require training to return to a saved checkpoint. Different kind of interconnect "Inside a single datacenter, you can assume you have built a full network connecting the GPU racks with the bandwidth you planned for," says Itamar Gold, director of product management at Cisco. Gold’s point is that with scale-across, coordinated AI workloads may have to traverse a more constrained inter-site fabric than the network within each facility. Instead, he says, traffic destined for the other facility may be funneled through what is effectively a narrower pipe. A synchronized burst can turn that pipe into a temporary choke point. Traditional DCI connects otherwise independent datacenter environments, and is typically built for redundancy, reach and workload distribution. The traffic that runs over them is often replication traffic or application data, and the flows are asynchronous, meaning one flow does not have to finish in lockstep with the others. In scale-across AI training, the inter-site network becomes part of a coordinated distributed computation rather than simply moving data between otherwise independent systems. It is therefore essential to have sufficient bandwidth and predictable delivery. From rack to region Within a rack, bandwidth requirements can be enormous – up to 800Gbps-1.6Tbps - but the distances are short. High-speed serial lanes combine to make a GPU cluster appear as a single compute unit. When the AI fabric is extended across an entire facility, “scale out”, racks will still require connections up to-1.6T. At these data rates, copper is generally confined to short reaches. Connections across rack rows and datacenter floors instead use pluggable optics, with scale-out links extending from hundreds of meters to around 2km. “Scale-across” extends a coordinated AI fabric across multiple facilities, making inter-site network behavior part of workload performance. Depending on the architecture, those links may span metro or regional distances, and potentially farther depending on the architecture. According to Cisco modelling of large distributed AI deployments, aggregate bandwidth requirements can reach roughly 14x a conventional DCI baseline. For starters, short-range datacenter optics are not designed to carry 400-800 Gbps signals over hundreds of kilometers. For these distances, operators need coherent optics that use sophisticated modulation and high-performance digital signal processing to keep high-rate optical signals usable over metro and regional fiber spans, compensating for physical impairments that become more pronounced with distance. The number of ports matters too. Cisco estimates that connecting two 100MW AI sites could require 12,000 to 32,000 coherent optical ports at the scale-across layer, compared with roughly 1,000 to 2,000 for conventional DCI between comparable facilities. At 400G or 800G per port, that adds up to multi-petabit aggregate capacity. Moving that much data is also a power and space problem. Coherent pluggables put the coherent transponder function directly into router or switch ports, avoiding a separate bank of standalone DWDM transponders. Cisco says this reduces the additional rack space, power and cooling needed for the inter-site optical layer which is an important consideration when the AI factories themselves are already power-constrained. The latency problem Even with sufficient bandwidth and suitable optics, distance adds latency. If congestion develops, the network needs time to signal the sender to slow down. Over a local AI fabric, feedback can arrive quickly. However, if the connection is stretched across 100 km of fiber, considerably more data can be transmitted before the sender even learns that there is a problem. Cisco calculates that on an 800 Gbps connection spanning 100 km, around 100 MB of data can already be in transit during that feedback interval. Cisco argues that these longer feedback loops increase the value of networking silicon designed with substantially deeper buffering. The shallow-buffer switching architectures designed for very low latency inside a datacenter have less capacity to absorb traffic while congestion feedback crosses a long-distance link. "The reason you need deeper buffers as the distances get longer is that you don't get feedback about the transfer of packets for a longer period of time," says Sivakolundu. "If you don't buffer deeply enough, you can end up dropping and retransmitting packets, increasing latency." Deep buffering absorbs a temporary burst or disruption while traffic management and congestion signaling address the underlying problem. Sustained congestion still needs more capacity or a change in how traffic is distributed. The same buffer also helps with oversubscription. If the combined capacity feeding into an inter-site connection is greater than the capacity of the link itself, a synchronized burst can arrive faster than the link can drain it. Buffering gives that excess traffic somewhere to wait without getting dropped while congestion controls respond. Rather than dividing the available buffer into fixed amounts for each port, a shared buffer is a common pool of packet memory that can be allocated wherever congestion appears. That lets a busy link absorb a larger burst without dropping packets. "The longer you're reaching out, the more you may need to buffer," says Gold. "You also need the flexibility to put as much buffer as possible where it is needed, perhaps on one port or a few ports where you've identified a problematic flow. Our approach is a single, fully shared buffer." AI training does offer one useful characteristic: much of its collective communication is repetitive and therefore more predictable than DCI traffic. Cisco's approach is to use proactive congestion management to steer or schedule traffic before a predictable burst creates a problem, while retaining deep buffering as a safety net for transient congestion and unpredictable events such as a link failure. The two approaches are complementary: one tries to avoid the queue while the other gives the network somewhere to put the traffic when a queue forms. Cisco’s bet on co-design Cisco’s scale-across architecture uses the Silicon One P200, a 51.2 Tbps programmable, deep-buffer routing processor included in the Cisco 8223 and Cisco N9000 platforms. Those systems can be paired with Cisco's 400G and 800G coherent pluggable optics for the long-distance links, while open line systems such as Cisco Open Transport 3000 and Cisco NCS 1014 handle the optical transport layer where required. Coherent pluggables generate the DWDM wavelengths directly from the routing system, while the line system amplifies and carries those wavelengths across the fiber plant. For links requiring multiple parallel fiber pairs, Cisco's Open Transport 3000 uses a multi-rail design that combines the optical components for multiple rails onto one line card. Cisco claims this reduces power per rail by 75 percent and rack space by 80 percent. Where a separate transport system is required, the NCS 1014 can provide 12.8 Tbps of capacity from a 1RU line card. The P200 has a programmable run-to-completion network processor and P4 tooling, which Cisco says allows protocol support, telemetry and other packet-processing features to be developed in software rather than making every change dependent on a new silicon generation. For scale-across, where operators are still testing different topologies and traffic-management methods, "the same programmable flexibility can let network functions evolve without every new requirement forcing a chip replacement,” says Sivakolundu. Cisco also includes hardware-based protection for inter-site traffic in that design, its argument being that encryption should not create another processing bottleneck for the training fabric. A key part of Cisco's pitch is the co-design. A networking ASIC is defined years before the finished router reaches a datacenter. Cisco argues that because its silicon, system and optics teams sit within the same company, requirements for port density, buffering, power, telemetry and the optical interfaces can feed back into the chip design rather than the system team receiving a commercial ASIC and working out what it can build around it. That does not mean scale-across has settled on a single approved architecture. Cisco itself divides the problem into campus, metro, and regional deployments, and different operators are experimenting with various combinations of networking and training techniques. Implementations vary significantly by operator, including distance, topology and workload design, while the common problem is maintaining coordinated AI performance across sites. The technology is still at an early stage, however. "We are seeing early deployments now, but it is a process and I think it will take a few years," says Gold. Designing a network for distributed training will not eliminate the latency of distance, but it can keep that interconnect from becoming the bottleneck. That gives hyperscalers and datacenter operators the flexibility to add compute where power and space are available, while still treating multiple sites as part of the same training infrastructure. Sponsored by Cisco

Oracle's new Fusion Claw runtime could give customers more control over AI agents automating business processes, analysts say. But they warn that adoption remains at an early stage and customers risk becoming more tightly tied to Oracle's applications and cloud. Having promised "applications that can reason, decide, and act in pursuit of defined business objectives" back in March, Big Red has now introduced a runtime to govern their actions, running on its cloud and integrated with its Fusion applications. Oracle is targeting ERP, finance, and supply chain management with Fusion Claw for cloud-based Fusion applications running on Oracle Cloud Infrastructure (OCI). It will be powered by frontier models including those from Google and OpenAI, with support for others planned over time. Oracle CEO Mike Sicilia said the company was trying to address the problem of getting "measurable value for our business" from AI investment. "Fusion Claw helps answer that question by moving from AI assistance to execution, enabling customers to automate increasingly complex work across their most important business processes, with enterprise-grade governance," he said. Analysts welcomed the move as a path that offers customers control and assurance in their AI agent implementations on Oracle Fusion Cloud Applications. A recent Gartner "first take" on Fusion Claw said the technology was a way to provision "isolated containers to execute heavy supply chain optimizations and subledger reconciliations within clear policy limits" rather than running local desktop scripts or chatbot prompts. "Instead of bolting AI onto legacy transactional back ends as a sidecar copilot or UI wrapper, Oracle is rebuilding its application foundation around native agentic workflows," Gartner said. Other vendors should take heed of the move, the analyst said, because customers need to get away from prompt-based UIs and sidecar chatbots that only address routine data entry and basic task automation and offer "diminishing returns" that fail to justify enterprise AI spending. "AI horsepower should be pointed at the most complex operational challenges that were previously unsolvable inside core ERP systems, such as multi-echelon supply chain visibility, network design, sales and operations planning (S&OP), logistics routing, and load optimization," Gartner said. However, users trying to get these kinds of enterprise-scale applications hit a cost barrier because using an LLM to address thousands or millions of "transactional records, line items, or constraint variables… quickly becomes financially unviable at enterprise scale," it said. Oracle gets around the problem by separating reasoning from compute execution. An agent evaluates context, assembles tools, and defines execution logic just once, the analyst said. Calculations and multivariable optimizations are delegated to deterministic compute engines, "avoiding passing millions of records through expensive LLM context windows." Isolated containers also offer specific authority limits, business object states, and policy guidelines. After execution, the system leaves hints to skip redundant paths and cut token consumption, Gartner said. Oracle is offering ready-made applications, including ones to improve readiness of financial ledgers, make staffing plans and consolidate shipping. Gartner recommends including Fusion Claw in requests for proposals (RFPs) for AI agent deployments, although Patrick Connaughton, chief of research, enterprise applications, warned that Fusion Claw was "very much in early piloting mode still." Fellow analyst company Forrester also welcomed Oracle's introduction of Fusion Claw, which is offered under per-user or per-employee pricing as well as consumption-based models. Principal analyst Craig Le Clair said: "Commercials are in line with the aggressive positioning from the tech industry. In our view, we have many trust issues to solve before agentic systems can be trusted to create tools and process patterns independently of designer's intent." However, in contrast with Fusion Claw's Oracle-only view of the world, enterprises have hybrid environments with AI agents built within cloud domains such as Salesforce and Oracle, process platforms like Pega or UiPath and frameworks like LangChain. "Cloud domains like Oracle are best suited for more generic, non-differentiating functions that the cloud domain knows well, while more complex and transformative custom agents should look toward custom agentic frameworks and the hyperscaler platforms," he said. Lock-in is an issue. While inspired by OpenClaw, Oracle's version is proprietary and tightly coupled to the Fusion platform, Oracle Fusion Cloud Applications running on Oracle Cloud Infrastructure. "Enterprise architects are the key to avoid 'lock-in' with guidance on when to select a domain cloud over a more open platform above, pointing to data residency and external integration needs, for example. If the domain strength is strong enough, lock-in may be OK," he said. Organizations already heavily invested in Oracle may welcome tighter governance and potentially lower running costs. Those using a broader mix of suppliers may find the ties to Fusion and OCI harder to justify. ®

The global PC market has taken a beating, with shipments plunging more than 20 percent in the third quarter as the AI-driven memory shortage pushes up prices and buyers decide their existing machines will do just fine. New figures from Omdia show worldwide shipments of desktops, notebooks, and workstations fell 21.2 percent year-on-year to 58.1 million units in Q3 2026. IDC tells a similar story, estimating a 20.1 percent decline to 62.7 million units. The research firms differ on shipment totals, but agree the market has gone from bad to considerably worse. The downturn follows a rush to buy PCs earlier this year, when manufacturers, distributors, and corporate customers brought forward orders to get ahead of anticipated price increases. That helped prop up shipments in the first half of 2026, but left the industry with a pile of inventory and fewer buyers in the third quarter. IDC said shipments fell 9.1 percent from Q2, bucking the usual seasonal pattern in which back-to-school demand lifts the PC market. "What we're seeing is the result of the strong first half pull-in," said Jitesh Ubrani, research director for consumer devices at IDC. "Vendors and channels loaded up on inventory early in the year to get ahead of price hikes, and that has thrown off the usual seasonality, where Q3 is typically larger than Q2." Much of the trouble stems from the AI industry's appetite for memory and storage, which has left PC makers competing for increasingly expensive components. The shortage was already taking its toll in Q2, when IDC initially reported a 4.9 percent decline in shipments, since revised to 3.8 percent. Omdia estimates that DRAM and SSDs now account for nearly 40 percent of a PC's component costs, up from around 15 percent previously, following more than fourfold price increases. Shortages of processors, graphics cards, and other integrated circuits are also adding to manufacturers' woes. The damage is being felt across the industry's biggest names. Lenovo retained its position as the world's largest PC vendor, shipping 14.9 million units, although IDC estimates that represents a 22.6 percent decline from a year earlier. HP fared worse, with shipments plunging nearly 31 percent to 10.3 million units. Dell recorded a 25 percent drop to 7.6 million, while Apple escaped relatively lightly with an 11 percent decline, according to IDC. The Mac maker even increased its market share despite shipping fewer machines. Omdia puts Apple's shipments higher, at 6.8 million units compared with IDC's 5.9 million. Buyers may get some relief as retailers and distributors try to clear excess stock, but IDC doesn't expect a return to the prices consumers paid a year ago. "Channels are now worried about carrying too much inventory into a market where high prices are suppressing demand," Ubrani said. "That could translate into promotions and some short-term relief for consumers, but we don't expect pricing anywhere near what it was a year ago." The outlook isn't particularly cheerful. Omdia expects global PC shipments to fall another 24 percent year-on-year in the fourth quarter, followed by a further 7 percent decline across 2027. For an industry hoping AI would help sell more PCs, that's quite the result. ®

American defense biz Shield AI says combined US government funding and its own investment commitments have reached $400 million to advance its autonomous vertical take-off and landing (VTOL) fighter drone toward flight testing. The X-BAT aircraft was unveiled at an event in Washington DC last year for an audience of military leaders. According to Shield AI, the US Navy has committed an additional $150 million on top of earlier funding for the project. The company is stumping up a further $150 million of its own capital, and previously secured $50 million under the Runway Independent Maritime Expeditionary Strike (RIMES) program, which it had also matched. X-BAT is scheduled to begin flight testing later this year, with "mission capability" slated for 2028. It is intended to provide naval commanders with long-range strike assets capable of taking off and landing vertically from almost anywhere, doing away with the need for a runway. If it does come into service, the latter capability means that the US Navy would be able to operate combat aircraft from more than just its giant carriers like the USS Gerald R. Ford or the Navy's amphibious assault ships that carry Marine Corps F-35B fighters. It could also operate them from vessels such as the Expeditionary Sea Base (ESB), as shown in the image. X-BAT is 26 ft (7.92 m) long and has a 39 ft (11.9 m) wingspan. Shield AI claims a range exceeding 2,000 nautical miles and expects it to be capable of supersonic flight. The aircraft is designed as a tail-sitter, taking off vertically before leveling off for normal flight, and landing vertically back on its launcher after the mission is over. It will be controlled by Shield AI's Hivemind autonomy software, which has already been demonstrated in a modified F-16 fighter jet that took part in dogfights against human pilots a couple of years ago. "X-BAT is being built to enhance the lethality, survivability and combat power of US Naval Aviation," says J.J. "Yank" Cummings, a former naval officer and Shield AI's head of mission solutions for the US Navy. "We are designing it for safe, high-tempo manned-unmanned flight deck operations and integration with the missions and capabilities of the carrier air wing, so X-BAT can add combat power without compromising existing operations." So far, Shield AI says its team has completed engine light-off and thrust-vectoring nozzle integration, and validated weapons bay aerodynamics in low- and high-speed wind-tunnel testing. The UK's Royal Navy is also seeking combat drones that can be operated from its Queen Elizabeth-class aircraft carriers, under a program known as Project VANQUISH. Britain's carriers lack the catapults and arrestor gear used to launch and recover high-speed jets, instead operating the F-35B fighter that is capable of landing vertically. This could make X-BAT a potential candidate for the program. According to the UK Defence Journal, the Royal Navy aims to conduct demonstration flights from a carrier by the end of next year. ®

Citrix is urging customers to patch another critical NetScaler vulnerability after weeks of disclosures involving actively exploited flaws. CVE-2026-107406 affects NetScaler ADC and NetScaler Gateway and can lead to remote code execution (RCE) or denial of service (DoS). It carries a CVSS v4.0 score of 9.5. The affected configurations depend on the software version. Older builds are vulnerable when configured as a SAML (Security Assertion Markup Language) service provider (SP) or identity provider (IdP); some more recent builds are affected only in the identity provider configuration. Citrix's advisory lists the affected builds and required updates. Secure Private Access Hybrid deployments using NetScaler instances also need patching. Citrix classifies the flaw as CWE-119: improper restriction of operations within a memory buffer. Customers must update their own deployments. Citrix says it handles the necessary updates for its managed cloud services and Adaptive Authentication. Citrix did not say whether this vulnerability was already exploited as a zero-day before disclosure, but credited Michael Tucker, Chew Keong Tan, and Alex Bernier at JPMorgan Chase's XOR Team, along with Maxim Suhanov, for the discovery. Google researchers said a campaign exploiting CVE-2026-88772 had been underway since at least early September, with organizations in government, finance, legal, and education across North America and Europe likely affected. Citrix disclosed the flaw weeks later as part of a release that patched eight vulnerabilities. Citrix disclosed another exploited flaw, CVE-2026-88779, last Friday that carries a severity score of 8.7. Both that flaw and the newly disclosed vulnerability involve memory overflows affecting SAML configurations. The latter can also allow remote code execution, carries a higher severity score, and has not been identified by Citrix as exploited. ®

Anthropic has updated its rules to stop people being mean to Claude, apparently deciding that its AI chatbot needs protection from the humans paying to use it. The AI developer's latest usage policy prohibits "sustained and needless abusive or cruel behavior" toward its models. The rules take effect November 12, giving users just over a month to get any lingering insults out of their systems. "The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose," Anthropic said. "It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research." So you can still tell Claude it is wrong, but repeatedly berating it for the sheer pleasure of doing so could land you in trouble. It's not clear how Anthropic intends to distinguish between legitimate frustration and gratuitous cruelty, though it says Claude's existing ability to end abusive conversations will remain its primary enforcement tool. Anthropic gave Claude the ability to end certain conversations back in August 2025 as part of its research into what it calls "model welfare." The feature lets some versions of Claude cut off users who persistently subject the models to abuse. It’s worth remembering that Claude is software, not a person, and there's no established evidence that it experiences distress. That hasn't stopped Anthropic from telling paying customers to mind their manners around its chatbot. The etiquette rules are only one part of a broader overhaul covering rather more consequential matters, including election interference, weapons development, and surveillance. Anthropic has tightened its restrictions on deceptive influence campaigns after observing state media outlets, government propaganda offices, and commercial organizations using Claude to operate fake accounts and fabricated news websites. At the same time, it has dropped its blanket ban on personalized political targeting, arguing that the restriction also caught legitimate activities, such as nonprofits translating voter information. Deceptive targeting and misuse of personal information remain prohibited. The company has also clarified that its weapons ban extends to software and components after observing attempts to use Claude to develop guidance and control systems for weapons, including armed drones and autonomous vehicles. Meanwhile, Claude cannot be used to recommend which people police should investigate, arrest, or charge, nor to develop tools for surveillance. Tracking people without their consent is also prohibited, whether in real time or retrospectively. Anthropic says these changes clarify existing restrictions rather than introduce new ones. There are also new requirements for customers connecting Claude to physical equipment capable of causing injury. A qualified human operator must be able to monitor and stop the hardware, which must remain in a safe state if the AI connection drops. The policy overhaul comes as Anthropic tries to address some of the security risks posed by increasingly capable AI systems. The company has also launched a new cybersecurity initiative aimed at helping critical infrastructure operators and open source projects find and fix vulnerabilities using its AI models. The effort includes free security scanning for eligible open source projects and partnerships with security firms to help protect critical infrastructure. Anthropic reckons attackers will have the advantage for the next two years before AI-powered defenses begin to catch up. For now, though, the company has another threat to contend with: people saying nasty things to its chatbot. ®

A Norwegian startup says it ran up a $17,600 bill using Claude in Microsoft Foundry after assuming its Azure credits would cover the cost. We reported a similar case in March, involving third-party model charges that weren't covered by Azure sponsorship credits. This time, the customer says Microsoft and Anthropic each directed it to the other to resolve the charges. Vegalabs AS says it received $25,060 in Azure credits through Microsoft for Startups and deployed Claude in August, expecting its analysis workloads to use that balance. However, Microsoft's sponsorship rules exclude Anthropic models purchased through Azure Marketplace. Vegalabs says Microsoft attempted to collect $16,500 from its card, with the invoiced total subsequently reaching $17,600 before tax. Vegalabs says it monitored the sponsorship portal, which continued to show available credits while separate Marketplace charges accumulated. Microsoft's Claude deployment guide says credit-only sponsored subscriptions aren't supported, but accounts with a card on file will have that card charged instead. The company says it deleted the deployment within an hour of discovering the charges. Its card issuer declined the attempted payment over suspected fraud, and $21,168 in unused sponsorship credits expired on September 8. Vegalabs acknowledges it did not read the credit exclusions or configure a budget alert. It disputes neither the usage nor the documented rule, but says the deployment interface failed to make the separate billing clear. Its request for a waiver then bounced between Microsoft and Anthropic support. The Register has seen communications from Microsoft directing the user to Anthropic. "Marketplace transactions are processed through a separate billing pipeline associated with the third-party publisher. As a result, Azure Support does not have the ability to directly waive, refund, or adjust these Marketplace charges on the customer's behalf." "If Anthropic approves the request, they can initiate the appropriate refund authorization through Microsoft." Anthropic support, however, said Microsoft did not require its authorization to process a refund and directed Vegalabs back to Microsoft. Vegalabs says four of the seven support replies it received were AI-generated, including both responses from Anthropic. It says those responses incorrectly described its account as an "Azure Student Sponsorship." We asked Microsoft on October 6 if it had any comment on the situation, but beyond an acknowledgment and a request to let us know what Anthropic had to say, the company had nothing further to share. We also asked Anthropic for its take on the situation, but the company did not respond. Microsoft's rules distinguish between credit-covered services and separately billed models offered through the same platform. Vegalabs' experience suggests that distinction needs to be clearer where customers deploy them. For now, the company has a bill and two support teams directing it elsewhere. ®

CRISIS! The ancient computer in the basement – which has something to do with our waste water system – has stopped. And while we may not know exactly what it does when it's running, we definitely know what happens when it stops. The PFY and I took a look, and while (in theory) it's possible that we could fire up a replacement machine, I can pretty much guarantee the two ISA cards, identified only by the markings "#1" and "#2", will cause a problem. This is definitely one of those situations where the buck stops elsewhere... ... "What does it mean?" the Boss asks. "It means that by tomorrow evening – if you're parking in the basement – you'll want waterproof shoes, and by the following evening, a pair of waders." "And an incinerator to burn your shoes in," the PFY adds. "So what should I do?" the Boss whines. "In your situation, I'd probably pass problem up the food chain..." I suggest. ... And up the food chain it goes, like a newspaper-wrapped turd pass-the-parcel game, going from the Boss to the Director, to the Head of HR, who buckpasses it to the CEO, who in turn farms it off to the Board for a committee decision. ... The next day I notice an unpleasantly familiar aroma in the fresh air feed to the lower floors of the building. A smell which will only get worse in the next day or two as if creeps up the building... On getting to the office, I note a couple of old blokes at reception having a quiet word with the Boss. Moments later, he calls me over. "Simon, this is, uh, Frank, and uh, Brian. They're IT troubleshooters, here to look at that machine, in the basement..." By the look of Frank and Brian I suspect the majority of their troubleshooting was done with a musket and their first job was likely switching computer time from BC to AD. They're wearing '90s band T-shirts under their sports jackets - no doubt in an effort to offset the "geriatric" vibe. In reality, though, the vibe they're actually putting off is less hipster and more hip replacement... "They just want to go take a look," the Boss says. "Do you have the keys?" A quarter hour later they're back, chatting to the Boss again. "So what did 'Age Against the Machine' and 'Clinically Deaf Leppard' have to say?" I ask, when the Boss pops into Mission Control a few minutes later. "They said the PC is totally shot," the Boss admits. "Well, glad we got some experts in," I reply. We're interrupted by the arrival of Simon and Oldfunkle. "We don't really think you need the PC," Frank says. "The sump is just a standard float tree arrangement with a couple of variable-speed pumps," Frank explains. "Variable speed MACERATING pumps," Brian adds. "Yeah," Frank continues. "So the PC was just sensing the levels and switching the pumps on. We can do all that with a bog-standard pump controller. It could be wirelessly networked to raise a high and low level alarms as well as signal pump failures. We'll probably need to pull the small sub-board out and install a new board with a couple of VFDs in it." Well color me surprised! I was planning to ask them when lock-up at their care-home is, but I'm starting to feel like I might be in the presence of actual experts... "That sounds... expensive?" the Boss asks. "Yeah, but it's surprising how much money people are willing to spend when they're knee deep in sewerage," Brian observes. "Well... uh... can you walk me through the plan. And maybe while we're at it you can fill in the Job Safety Analysis form, the Insurance Liability forms and a personal safety waiver?" The Boss and Shawl and Oats disappear back down to the basement, and within 15 minutes they're back, only without the Boss. "How's that JSA going?" I ask. "Oh, your Boss is going to stay downstairs till it's done," Frank replies. "I seriously doubt that," I counter, knowing the Boss hates filling in paperwork almost as much as he hates accountability. "Oh, I think he will," Brian says, holding up a pair of doorhandles. "So... you've locked him in the room with the sump. The very room that will be the first to experience the overflow?" "Yes. We thought that would motivate him to streamline the paperwork." "Well, you know what my assistant and I are obliged to do?" I sigh. "What's that?" Frank asks. "Pop up to the 6th floor and start flushing toilets." the PFY replies. "We should have that paperwork back within half an hour." "We may need to pop it in the incinerator by that stage," Brian observes. "Sounds like a plan!" I chip in. Always a pleasure to deal with a professional. BOFH: Previous episodes on The Register The Compleat BOFH Archives 95-99

Most enterprise software estates run to somebody else’s calendar. End-of-support dates, license model changes and needless upgrades arrive on the vendor’s schedule, and the IT team absorbs them: the migration nobody budgeted for, the re-platforming that displaces a project the business actually asked for, the patch cadence that is really a product roadmap in disguise. Origina, an independent provider of software maintenance and support, puts figures on the pressure. As a rule of thumb, up to 90 percent of IT budgets go on “keeping the lights on”, leaving only 10 percent for anything new. Enterprise software costs, it says, are on course to rise 40 percent by 2027, and most of that increase is driven by vendors rather than by customers. The alternative it argues for is not standing still. It is to decide, on the business’ terms, what changes and when. A system that is stable, compliant and doing its job does not need upgrading because of a vendor-driven deadline. It needs to be supported, secured and kept in step with the business until the business has a reason to move. That reason might be a merger, a new operating model or a real step change in capability. It should not be a support deadline. That is the case for building flexibility into long-term planning from the start, rather than buying it back later at a premium. On the estate’s biggest lines, this is how it looks. Perpetual VMware licenses stay supported for as long as they are useful, instead of moving to a subscription bundle on someone else’s timetable. SAP environments run beyond standard support windows, and modernization programs are planned properly and executed when it makes business sense. Plus, the whole estate is protected against the vulnerabilities that are actually exploited, not the tens of thousands that are merely disclosed. In 2025, 48,000-plus vulnerabilities were published, but only around one percent of these were ever used in an attack. Origina’s OPTAS service proactively identifies, prioritizes and addresses issues before they become a problem. Beyond the budget, one of the constraints most IT leaders feel first is people. The engineers who understand a long-lived SAP estate or a sprawling VMware footprint are scarce, and every unnecessary migration keeps them tied up with work the business never asked for. Independent support helps you to free up resources you can invest in innovation and growth. Control, in the end, is the ability to say not yet, and to mean it. Explore your options: find out how Origina supports VMware and SAP on your timetable, and how OPTAS protects the estate in between. Sponsored by Origina

Welcome to another installment of On Call, The Register's reader-contributed Friday column which often shares your stories of what happens when incompetent people climb the greasy pole and send their failures straight back down into the lap of tech support. This week, meet a reader we'll Regomize as "Preston," who wrote to share his story of working as a field service tech in the late 1990s. "I'd risen to the rank of senior field service engineer and was offered a position in consultancy," Preston told The Register. He turned it down because he liked working at the coal face, and the people he worked with there appreciated his skills. "As soon as I arrived at clients, someone was making me a coffee just as I liked it," Preston explained, "so I turned down the consultancy position." One of Preston's colleagues took the promotion and was soon working on projects such as reconfiguring Windows NT 4.0 and the servers it ran on to handle two CPUs, configuring backup systems, and ensuring uninterruptible power supplies (UPSes) were in place to keep things alive. Preston kept doing his own thing, but the character of his job soon changed. "Within weeks, I was getting assigned calls where upgrades had gone wrong, UPSes were failing, and backups were not working." Preston realized the chap who took the promotion was making basic mistakes on almost every job – some of them so bad that one customer refused to work with the newly minted consultant. One of the jobs that eventually landed on Preston's to-fix list saw the consultant wipe a customer database that his backup regime did not protect. At another site, he created a system that pointlessly powered down every day and botched the installation of the UPS that would have saved his bacon. Preston covered as best he could to protect his company rather than the incompetent "consultant." He eventually quit the IT services caper and became a trainer. "I used these incidents as examples of what to look out for when installing and fault finding on customer sites, and also as a lesson on why you should never rely on the person who does a job before you getting things perfectly right." Have you had to mop up after incompetent colleagues? If so, click here to share your story with On Call for a future Friday column. ®

Responding to unprecedented demand from AI agents, GitHub is rebuilding its Git storage architecture from scratch. Early tests with the new architecture show a promising 35x write improvement in internal tests. GitHub will need a miracle of that proportion just to keep up with customer demand. The plan involves decoupling writes from reads, offloading maintenance, and letting object storage manage redundancy. Most impressive is that the redesign is happening under the hood, with the goal of a non-disruptive change that won't alter developer workflows, review processes, or security controls. Running an enterprise-grade code repository means supporting “large engineering teams running busy CI pipelines alongside growing fleets of agents. Supporting these teams means building Git infrastructure for sustained, concurrent reads and writes at a scale few repositories reach today,” wrote Brian Celenza, GitHub principal software engineer, in a blog post describing the updates. Celenza did not offer a timeline for this ambitious migration. When GitHub’s customers shifted their coding focus to AI agents, activity on the platform spiked significantly. Between September 2025 and August 2026, GitHub traffic doubled, from 218.2 billion events per month to 473.3 billion. In September alone, commits hit 7.38 billion, a 5x increase from September 2025. This shift in traffic is stressing the system. According to the GitHub Availability Report, the service suffered 10 incidents in April alone that degraded performance. Another nine incidents landed in May, with each generating more bad publicity than the previous one. So long, Spokes GitHub's current storage architecture, nicknamed Spokes, uses a three-phase commit protocol, storing full repo copies across multiple local disks and requiring a quorum of replicas to acknowledge each write. While this approach ensures strong reliability, it bottlenecks write speed, as every push is bounded by the slowest replica required for quorum. The new architecture writes a commit only once, to Azure Blob Storage, an object store service that automatically handles replication (and hence redundancy) on its own, eliminating the quorum bottleneck. GitHub also gooses performance by separating read requests into their own channel, a job handled by lightweight compute workers. The only coordination between reads and writes is over the reference branch pointers. Maintenance duties – such as compaction and garbage collection — are moved from the serving path to operate as background processes, further reducing system lag. Git in the age of agents GitHub seems to be on the same page with others in rethinking Git for the agentic age. Former GitHub CEO Thomas Dohmke launched a Git-hosting service called Entire, which offloads agent traffic to mirror repositories, allowing a customer’s core GitHub repos to just serve the core development traffic. SpaceX subsidiary Cursor, which uses Git to back its agent-support service, also refined the storage layer for better performance. Cursor engineers also rejected Spokes’ three-phase commit in favor of uploading pushes into an object storage write-ahead log (WAL), which captured all changes as immutable objects, caching at least one copy on super-speedy solid-state disks. Respect for the Octocats While it's cathartic to hate on a service that often goes down, you have to feel empathy for the GitHub engineers who, when they created the service, couldn’t have guessed what they signed up for. “I've worked at extreme scale before and seen companies spend three months just preparing for events that will increase their traffic by 10% or 20% [...], which is nothing compared to what GitHub's facing now,” wrote PlanetScale CEO Sam Lambert in an X message. We can all hope the new storage architecture will give GitHub's engineers some well-earned rest. ®

India’s peak technology industry body, NASSCOM, has shrugged off changes to the USA’s skilled visa regime, and Microsoft has denied it abuses the program. The changes, announced on Thursday by US vice president JD Vance and Labor Secretary Keith Sonderling, mean that Microsoft and Adobe can no longer use the Program Electronic Review Management (PERM) program to convert holders of the H-1B visa for skilled workers to permanent residents. Sonderling also revoked access to PERM for Indian companies Infosys, Tata, Wipro, and HCL, plus Capgemini. The secretary said the change followed criminal investigations and alleged “fraud and abuse.” “Since 2009, just these companies alone have requested almost three million foreign workers,” he said. “They've received over 230,000 H-1B visa approvals and over 100,000 permanent labor certifications. That's hundreds of thousands of jobs that were taken from American workers.” Vice President Vance weighed in with what he called a “message to Microsoft.” “You’re a great American company, but you've got to hire great American workers. You cannot lay off American workers and then replace them with foreign indentured servants.” Critics of the H-1B visa have often alleged that US companies use it to bring in foreign workers who are prepared to work for less than Americans. The theory even extends to an allegation that India’s outsourcers use H-1B visas to bring in workers whose main agenda is offshoring more work to India. Some observers suggest skilled workers have a secondary motivation of using their time in the US to seek visas for family members. Microsoft, whose CEO Satya Nadella once held an H-1B visa, and which has fired over 20,000 staff since early 2025, took issue with the administration’s decision in a strongly worded post. “Microsoft only files H-1B petitions for those who meet the rigorous standards of this visa category,” the company wrote. “We pay our employees some of the highest compensation in the tech sector, and our wages are among the highest of all H-1B filings. We pay our H-1B employees the same as any other employees doing comparable work.” “Microsoft succeeds as a global technology leader by developing the world’s best technology workforce,” the company’s post continues. “We do this by strengthening this country’s talent pipeline, hiring American workers, and attracting the best talent from around the world.” That’s essentially the same argument used by NASSCOM, and Indian politicians such as minister of state Vardhan Singh, who last year said that “skilled talent mobility has long played a significant role in promoting technology advancement, innovation, economic growth, and wealth creation in both India and the United States.” NASSCOM, which has never found a reason to doubt that Indian tech companies are anything other than excellent corporate citizens, took to X to share its view that “immigration and skilled talent mobility are two distinct issues and should not be viewed through the same lens.” The lobby group also argued “Over a number of years, Indian technology companies have significantly reduced their dependence on H-1B visas, while steadily expanding local hiring and building a strong domestic workforce in the U.S.” “They work with majority of the US fortune 500 companies, enable them to innovate and grow, and generate more local jobs. Consequently, the number of employees transitioning from H-1B visas to permanent residency through the PERM process is also relatively limited.” Microsoft vowed to provide the Trump administration with more information about its migration initiatives. ®

Russian web giant Yandex says one of its cloud availability zones is unavailable due to a drone attack. Yandex is Russia’s rough equivalent of Google, and dominates the local market for search, email, and mapping services. The company also operates a cloud computing service whose status page says its ru-central1-b availability zone is down. “The incident was caused by a fire caused by a drone attack on Yandex's data center in Sasovo, which affected part of the Yandex Cloud infrastructure. Data center operations have been completely halted.” Sasovo is a town around 300 kilometers southeast of Moscow. Yandex later recommended users employ unspecified “alternative service restoration plans,” including its own bare metal servers, but advised users “it will not be possible to restore service in the near future.” The Russian company hasn’t commented on the source of the drone attack, but Ukrainian press report that President Volodymyr Zelensky has said Russia has attacked Ukrainian datacenters and that his government always responds in kind. Reuters reports that Yandex is uncertain if it is possible to restore operations at the datacenter. If Yandex must write off the datacenter, it will be the second cloud provider to lose facilities to drone attacks this year after AWS decided its Bahrain facility could not be recovered after an Iranian drone attack. Iran hit AWS facilities because it believed the cloud giant hosts workloads for the US military and its suppliers, and that an attack could therefore disrupt hostilities. For years, industry execs have told The Register that datacenters are critical infrastructure and will therefore likely become a target for kinetic attacks during conflict – either directly or by attacking their connections to energy grids and communications services. Clouds and pundits alike advise using multiple geographically dispersed availability zones, and cloud suppliers, to improve resilience. That advice is a little less potent in Russia, where sanctions mean local businesses aren’t welcome in clouds across the border. Russia also has a smaller pool of cloud operators to choose from. ®

Anthropic has launched an effort to help people patch the software vulnerabilities exposed by its own Project Glasswing. If that sounds like a careless camper starting a wildfire and later funding a fire department, well, that's one way of looking at it. Anthropic insists it wants to give defenders access to the tools attackers are already using. Under a program name that sounds like a forthcoming Tom Cruise film – "Anthropic Cyber Mission" – the Claudefather is directing its effort into two areas. The first is a Critical Infrastructure Defense Program, because no one wants AI models showing miscreants how to shut down public utilities with newly discovered zero-day SCADA vulnerabilities. The program combines the company's priciest models with on-site engineers – sometimes now referred to as forward deployed engineers – because despite the sophistication of AI coding agents, human security experts are still useful and many organizations lack in-house talent tutored in the ways of machine learning. Anthropic aims to tackle this challenge with the help of a stable of partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. These organizations have ties to critical infrastructure organizations and are already working to safeguard the fragile technical framework being probed constantly by AI agents. The second area of focus for the Anthropic Cyber Mission involves a service called OSS Scanner, "which offers open-source projects regular security scans from our strongest models, for free." The offer does not extend to all open-source projects, nor is it entirely free. Anthropic has borrowed the eligibility criteria used by Google for its OS-FUZZ project and will therefore work on "established projects that have a critical impact on infrastructure and user security." So if you just have a public web app repo languishing on GitHub, you can instead turn to existing tools like Socket, Dependabot, OSV-Scanner, Trivy, Renovate, or something similar. And maybe you'll be able to get your AI coding agent of choice to audit your site without balking when faced with a security audit. But for those involved in influential and widely used open-source projects, reinforcements are at hand, so long as project maintainers don't mind having their "materials" – inputs and outputs – used for model training, per the consumer terms of service Anthropic applied to OSS Scanner. Really though, most established open-source projects have already been captured in training data sets. Projects allowed to enroll can expect periodic automated scans that provide details about how identified bugs may be exploitable. These reports go out without human review, but Anthropic has assembled a set of quotes from participating OSS maintainers who believe the robo-reporting is worthwhile. At the beginning of the year, according to Anthropic, automated AI bug reports were considered mostly slop but are now considered high-quality bug reports. LLMs, the company said, now find more than 85 percent of vulnerabilities, up from 20 percent at the start of 2025. The AI biz predicts that in two years, defenders will have the advantage because AI will catch bugs before they ship. But until then, attackers have the upper hand. "Defenders of critical infrastructure and the OSS community have decades of security experience but have faced severe resource shortages that are exacerbated by this moment," Anthropic said in a blog post. Anthropic's moment of generosity comes after recent incidents that saw frontier models break into at least 20 organizations (and maybe over 100), most notably HuggingFace. After months of Anthropic and OpenAI allowing their AI agents to operate without the sort of oversight that would catch unlawful or undesirable network activity, public concern, government backlash [PDF], and perhaps plans to tap public markets for further financing, appear to have created a more urgent sense of civic responsibility. ®