The Register

The FBI announced that it has seized seven web domains linked to hacking tools allegedly operated by a Chinese security firm called Integrity Technology Group and used by Beijing-backed cyber operatives to scan a South Carolina power company's network and other critical infrastructure systems for vulnerabilities. In a subsequent advisory, the FBI and other government agencies in the US, UK, Australia, Canada, Japan, New Zealand, and Spain warned that Chinese government-linked attackers, enabled by Integrity Tech, are using botnets, malware, and other intrusion tools to target organizations worldwide and steal sensitive data, including from US critical infrastructure networks. “These actors exploit vulnerabilities by using scanning tools, cross-site scripting attacks, and password spraying on Microsoft Exchange servers, while establishing persistence through VPN software and exfiltrating emails and credentials using scripts,” according to the security alert. Based on this activity, the US Cybersecurity and Infrastructure Security Agency (CISA) has added five CVEs to its Known Exploited Vulnerabilities Catalog: CVE-2015-3306 CVE-2015-5477 CVE-2016-3081 CVE-2021-3199 CVE-2023-22894 The court-authorized seizures are the latest in a long series of US law-enforcement attempts to disrupt a Beijing-backed cybercrew called Flax Typhoon and shut down its botnet. From 2021 until its disruption, Flax Typhoon allegedly used a version of this Mirai-based botnet to infect internet-connected devices with malware, scan networks for vulnerabilities, and launch additional cyberattacks, all while hiding the PRC government hackers’ true IP addresses and physical location. The feds allege Integrity Tech developed the botnet and a vulnerability scanner called Microscan, and operated a post-compromise tool called FishHub. The latter allegedly downloaded additional malware to the phishing victims’ networks and stole sensitive data. Court documents, unsealed on Thursday, allege Integrity Tech has contracts with the PRC government, and the feds have long linked Flax Typhoon to the private firm. “Flax Typhoon actors conducted successful computer intrusions against multiple victim entities which had been scanned using the Microscan tool,” according to the court documents. Victims include a university in Hsinchu, Taiwan, that Flax Typhoon compromised in March 2023, and a second university in Puli Township, Taiwan breached in August 2022. “On or about April 26, 2022, and on or about December 29, 2022, Flax Typhoon actors also used the Microscan tool to scan for vulnerabilities on the networks of a U.S. power company based in South Carolina, a multi-national Non-Governmental Organization, Japanese and Polish airports, and at least two Taiwanese critical infrastructure companies in the natural gas and power sectors,” the documents allege. Integrity Tech accessed Microscan through one of the seized domains, c0cc[.]cc. Meanwhile, five of the seized domains, 98aicai[.]com, 98aicode[.]com, outlook3650[.]com, youtubecard[.]com, and linkedinns[.]net, delivered the FishHub malware as recently as March, and infected about 20 Taiwanese universities. “Based on my training and experience I believe the tool was named FishHub because it facilitated phishing activity,” FBI special agent Adam James said in a seizure warrant affidavit. After FishHub downloaded a file onto victims' computers, that file used the five domains to retrieve additional malicious programs and code. The malware created a file listing, searched for specific files, compressed documents, and then exfiltrated the selected files to an attacker-controlled server. This gave Flax Typhoon remote access to compromised networks. In September 2024, the FBI said Integrity Tech and Flax Typhoon tore down their 260,000-device botnet after the FBI and its international pals went after them. As recently as this year, however, both private sector security researchers and Western governments warned that Chinese hackers continue to turn compromised routers and IoT devices into botnets and break into critical networks. In February, operational technology (OT) security provider Dragos said China’s state-sponsored spies haven't let up on their attempts to compromise America's critical infrastructure, and this includes a group whose activity overlaps with Flax Typhoon. This group focuses on gaining long-term access to OT engineering workstations and exfiltrating operational files, and it targets manufacturing, defense, automotive, electric power, oil and gas, and government organizations across the US, Europe, and the Asia-Pacific region. In April, a 10-country joint advisory warned that basically every Chinese “Typhoon” group is using botnets “strategically, and at scale.” Some of these covert networks are built and maintained by Chinese infosec companies including Integrity Tech, which also controlled and managed the so-called Raptor Train network, the alert said. In late August, the FBI said that it had disrupted a different botnet and seized domains associated with two platforms that Chinese government-backed operatives allegedly used to target NASA, the US Senate, the Department of Energy, and several other government agencies and critical networks. The DOJ subsequently clarified that not all the named organizations had been compromised.®

Nvidia on Thursday scraped together some spare change to commit $1 billion — about 1/60th of its quarterly profits — to funding US scientific discovery over the next five years. The commitment, announced during an event in Washington, DC, aims to support the research and development of AI — or is it "super intelligence" now? — in fields including quantum computing, healthcare, and energy security. Nvidia's choice of technologies is not surprising as it has long aspired to fuse AI and high-performance computing. And while LLM training and inference pay Nvidia's bills, the GPU giant has continued to introduce new CUDA libraries specifically for AI-assisted quantum computing, healthcare, drug discovery, and physics simulation. The initiative is part of the US government's broader Genesis Mission, announced late last year by the Trump administration, which aims to use AI to drive the scientific discoveries necessary to ensure the country's continued technological leadership. The program marks Nvidia's return to US supercomputing in a big way. The Department of Energy's (DoE) last flagship supercomputers powered by Nvidia, Summitand Sierra, were commissioned all the way back in 2018. The A100-based Perlmutter system launched in 2021 is also notable, but fell far short of the older systems. And while Nvidia never stopped building supers for the US government, they tended to be significantly smaller than the massive AMD-based platforms like Frontier and El Capitan. That changed last year when Nvidia revealed it would supply the US government with no fewer than seven new supercomputers across the Argonne, Los Alamos, and Lawrence Berkeley National Laboratories. Argonne will house the 100,000 Blackwell-GPU-based Solstice system. Built in collaboration with Oracle, the machine is largely aimed at emerging AI workloads, which don't depend on ultra-high precision computation and instead can get by with 16, eight or even four bits of precision. That puts Nvidia in a strong position to meet the US government's goals under the Genesis Mission. Over the past few generations of GPUs, Nvidia has traded much of its accelerators' double-precision FP64 grunt in favor of higher throughput at precisions used by more lucrative AI workloads. In the case of its upcoming Vera Rubin platform, Nvidia relies heavily on an FP64 emulation capability based on the Ozaki scheme we looked at in more detail earlier this year. Los Alamos' upcoming Mission and Vision systems will be powered by these same chips and take advantage of Nvidia's Quantum-X800 InfiniBand. FP64 emulation comes with compromises, but for workloads that are predominantly AI-optimized and only occasionally need higher precision, it can be worthwhile for the additional matrix FLOPS. While Nvidia is playing a much bigger role in US scientific computing efforts, it's worth noting that the DoE's biggest super — at least for pure FP64 — will still be powered by AMD. Oak Ridge National Laboratory's Discovery system, which is expected to come online 2029 will be built by HPE's Cray division and feature The House of Zen's FP64-optimized MI430X GPUs and Venice CPUs. Discovery is expected to offer peak theoretical performance somewhere between 3.3 and 8.5 exaFLOPS depending on whether or not Oak Ridge gets a facility power upgrade. Meanwhile, China's LineShine leads the Top500 ranking of Earth's mightiest computers. That system made its debut this northern spring with 2.2 exaFLOPS measured out of a 2.7 exaFLOPS of theoretical performance. With Uncle Sam's nose bloodied by China's return to the ranking, Oak Ridge may not have to fight hard for the extra power. ®

Google Cloud CEO Thomas Kurian said that Gemini will be "your new single, universal agent for work," a declaration as unassailable as Google's history of consistency, customer service, and long-term product support. Long-term here means a median lifespan of about 4.1 years and a mean lifespan of about 5.4 years across 308 discontinued products and services. That doesn't include enduring successes like Gmail and Search. Perhaps Gemini is destined for that level of longevity. In a blog post adapted from the prepared remarks he spoke at the Gemini at Work 2026 event – and what could be more appropriate for Google than to herald Gemini with a derivative work? – Kurian laid out his vision for an AI service known as Bard only two years ago. Gemini, he said, has your business context – which is to say, data – and thus can be used for various tasks like question answering, content creation, and coding. "It plans the work, uses skills and tools, connects to your systems, and brings back something finished — inside the documents, the inbox, and the developer environments you already work in," Kurian said. "It chooses the best model for the job, has built-in cost controls, and most importantly, has the security, administration, and governance required by your company." Kurian's actual remarks at the event were more succinct. "Today, Gemini became an agent," he said. "You give it objectives, not just instructions. It plans the work, uses custom skills and tools, connects to your systems and brings back something finished." You might wonder what role you play in all this as a Google Enterprise customer, when Gemini is planning, executing, and delivering. There's always Minesweeper and hunting for a job that involves physical labor and human dignity. Sundai Pichai, CEO of Google parent Alphabet, opened the Google Cloud event with the Recitation of Self-Reported Statistics, a common ritual in the technocratic liturgy. "Enterprise adoption is moving faster than most people realize," he said. "Nearly 80 percent of Google Cloud customers are using our AI products. And they are not just using it, they're using it intensely. Nearly 500 enterprise customers have each processed over 1 trillion tokens on our platform in the last year." Pichai took time to praise Google's latest model, Gemini 4 Argon, noting that it outperforms rival models Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra on the DeepSWE v1.1 coding benchmark with a score of 77.9 percent. And within Google, he said, Gemini 4 Argon has reduced post-submit code rollbacks by 30 percent while increasing agentic code submissions from employees by 35 percent. Kurian then continued the customer success report by citing various claimed productivity gains from Google Cloud customers. Japanese telecom biz NTT Docomo, he said, has saved 450,000 hours annually using data agents. Your mileage may vary. But Kurian's emphasis on the solipsism of Gemini isn't entirely accurate. Gemini is unified in theory. It works within Gmail, Dive, Docs, Slides, Sheets, Chat, and Calendar as a single surface. It can be accessed across iOS, Android, Windows, and Mac desktops, not to mention third-party applications where enabled or during headless operations. It resides in Google Cloud so it retains contextual info. But no agent is an island. Just like the stellar twins for which it's named, Gemini has more than one moving part. "It can dynamically create a roster of sub-agents — temporary, job-specific agents, each with their own identity — to tackle multi-step tasks," said Kurian, perhaps overestimating the appeal of naming spawned processes. And the Gemini agent is distinct from the various models it may call upon, even if one of those might be Gemini 4 Argon. "It runs each job on the model that fits best — orchestrating across our Gemini family of models and Claude models from Anthropic today, and other leading private and open models in the future — to deliver optimal quality and lower your costs," Kurian explained. During the event, Kurian highlighted how Gemini inherits existing enterprise controls. "Every Gemini agent has an identity of its own, separate from your identity," he said. "Your policies reach it the very same way they reach your people: set the policies once, apply them everywhere rather than agency by agent. You set its budget before it starts and that budget covers the whole cost of the work." That's the promise, but of course the very existence of enterprise cost control mechanisms suggests costs can get out of hand. It was only two months ago that Google introduced flexible billing and cost controls for agents. Managing AI spending, however, remains an unsolved problem. As noted by the Wall Street Journal recently, "Spending on AI Is Becoming Almost Impossible for Businesses to Budget." Back in July, AI cost management company Coworker AI observed there were two Gemini Enterprise products that could easily be confused. Now we're told there's just one. Enjoy it while it lasts. ®

NASA and the Department of Energy have agreed to expand their collaboration on the space agency's nuclear ambitions, as administrator Jared Isaacman pushes ahead with a tight deadline to launch a fission-powered spacecraft to Mars. “We are entering the ‘Nuclear NASA-era,’” Isaacman said at a signing ceremony where he and Energy Secretary Chris Wright signed a joint memorandum of understanding. The MOU, the full text of which doesn’t appear to have been published, establishes a framework for collaboration on space nuclear technology. Per NASA, the agreement “unites both agencies” in pursuing its objectives, with work to include research and fuel production, testing, launch integration, and operations. “Nuclear power will allow us to go farther, operate longer, and field more capable spacecraft and instruments than ever before,” Isaacman added. “The work we’re doing today is laying the foundation for the fission-poweuing NASA’s objred spacecraft of tomorrow and opening an entirely new frontier for exploration and discovery.” NASA has been making big plans to launch nuclear reactors into space since earlier this year when the agency paused work on the orbital Lunar Gateway station in favor of a base on the Moon’s surface. Those plans include a nuclear reactor to power said base, which NASA said on Thursday it intends to have ready for launch by 2030, giving the agency precious little time to meet such an ambitious goal. NASA has provided little evidence the initiative is on track, aside from noting in Thursday's announcement that it is “making steady progress” on Lunar Reactor-1 (LR-1), the unit it wants to use to power its Moon base. Before then, however, NASA faces an even tighter deadline for Space Reactor-1 (SR-1), which it intends to use to power a nuclear-propelled spacecraft scheduled to launch in December 2028. That’s just over two years away, mind you, and NASA has provided little public evidence of progress on the spacecraft, which it has dubbed "Freedom." NASA provided details [PDF] for SR-1 back in March when it paused work on the Lunar Gateway, describing the spacecraft as being mostly constructed from a repurposed spacecraft bus that NASA had almost completed for a previous project. The reactor would produce more than 20 kilowatts of electricity to power Freedom's electric thrusters on its journey to Mars, where it would deliver three Ingenuity-class helicopters to survey potential future human landing sites. Whether that’ll happen on any sort of timeline that resembles the current one is anyone’s guess, but let’s face it: NASA doesn’t exactly have a great track record for getting things done on time. The Government Accountability Office has been watching NASA’s lunar return ambitions for years, and has been noting for nearly a decade that its timelines were more fantasy than reality. In the past couple of years, the GAO has noted that existing NASA timelines were tight enough that any one mishap could push them back considerably, and massive workforce reductions under the second Trump administration haven’t helped matters. The Artemis missions to return humans to the Moon have also suffered a few hitches. Artemis I and Artemis II faced multiple delays, and NASA has since changed Artemis III's objectives, turning the planned 2027 lunar landing into a demonstration mission in low Earth orbit, in part to address concerns about the readiness of lunar landing systems, including SpaceX's Starship. We need not remind anyone of the saga of Boeing's calamity capsule. Who is going to launch that nuclear reactor in just over two years' time, and will NASA be able to land humans reliably on the Moon's surface by then? We checked in with NASA to get some assurances things were actually on pace, but didn’t hear back. ®

Researchers found thousands of GPU servers exposing Nvidia's DCGM Exporter to the internet, with hundreds potentially vulnerable to a high-severity flaw that could let unauthenticated attackers crash the GPU monitoring service and disrupt AI workloads. DCGM Exporters read telemetry from the GPUs on a host, including its hardware, utilization, memory usage, power consumption, and error events. Each GPU has its own unique ID, or UUID, and all of these metrics are exposed in plaintext over HTTP. This exposure provides would-be attackers with detailed information useful for reconnaissance, including mapping GPU infrastructure, identifying potentially vulnerable systems, and monitoring workload activity. Michael Katchinskiy, a researcher at datacenter security startup Lava, found and reported the bug in the GPU health and performance monitoring service. In September, the GPU giant Nvidia released a fix for the flaw, tracked as CVE-2026-47483, and gave it an 8.2 CVSS high-severity rating. “Once we realized how much these endpoints revealed, the next question was: How many of them are exposed to the internet?” Katchinskiy said in a Thursday blog. So the researchers started scanning the internet for exposed DCGM Exporters. And the scale of exposure proved “especially significant,” he wrote. Over the course of four scans between March and May, the threat hunters found about 2,100 GPU servers exposing DCGM Exporter metrics to the open internet. These included 12,000 GPU UUIDs. None of these required authentication. The hosts belonged to about 300 organizations, according to Katchinskiy, and nearly half - 5,274 of the exposed GPUs, or 44 percent of the total - were located in the US. These GPUs represented about $100 million in hardware, and included Nvidia Blackwell Ultra B300 GPUs, H200s, and H100s - used to run large-scale AI workloads - plus consumer RTX 5090 and 4090 systems. While investigating the exposed systems, the Lava team found that about 25 percent of the exposed DCGM hosts also revealed data from Go’s /debug/pprof/ built-in profiling tool. The profiler collects and exposes runtime performance data such as CPU and memory usage for running Go applications. This includes CPU usage, memory allocations, goroutine states, and blocking events. “With enough concurrent unauthenticated requests, the exporter could run out of memory and crash, cutting off visibility into GPU health and activity,” Katchinskiy wrote. The CPU and memory pressure could also affect AI training or inference workloads. Nvidia fixed the issue in version 4.8.2, and operators should upgrade to that version or later. In addition to GPU telemetry, Lava looked into Prometheus Node Exporter, which monitors server hardware and operating systems, and also exposes metrics over HTTP. The team found 12,096 public Node Exporter hosts exposing data on server models, operating systems, firmware versions, hostnames, storage paths and networking hardware commonly used in GPU clusters. This information reveals how environments are built and configured, which could also be used by attackers for reconnaissance, matching the system to known vulnerabilities. The publicly exposed monitoring services affected customer infrastructure across neocloud and GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data, and DigitalOcean. Lava reported all of this to the affected providers, and Katchinskiy says that these providers worked with customers to address the exposures. For operators: the security shop says Nvidia DCGM Exporter, Node Exporter and Prometheus services should not be directly reachable from the public internet and recommends restricting them to authorized monitoring infrastructure. “The findings highlight a growing security gap in AI infrastructure: companies are spending millions on GPUs while leaving critical systems exposed,” Katchinskiy wrote. “Those exposures can reveal how AI environments are built and, in some cases, allow attackers to disrupt them.” ®

A hole in America's domestic supply chain is closing. US-based wafer-fab GlobalFoundries announced a $2 billion partnership with TSMC on Thursday to shore up American manufacturing of advanced semiconductors, like those used in AI datacenters. Under the multi-year deal, GloFo will produce silicon interposers for TSMC's chip-on-wafer-on-substrate (CoWoS) advanced packaging spec at its Malta, New York fab site. Silicon interposers are essential to the production of many multi-die chip assemblies like GPUs and AI accelerators. These interposers sit under the compute logic and memory and provide a higher speed and lower power path for communications between the chips than would be possible using conventional organic packaging. Basically every chip that employs high-bandwidth memory (HBM) today relies on some form of interposer or interconnect bridge to stitch everything together. GlobalFoundries' partnership with TSMC appears to target the former. Advanced packaging has become a major bottleneck for American chip manufacturing. Notably, when TSMC began producing Nvidia's Blackwell generation of GPUs in its Arizona chip plant last year, we noted that the final product would still need to be round tripped to Taiwan and back for final assembly. TSMC has contracted with Amkor, an outsourced semiconductor assembly and test (OSAT) provider, for CoWoS compatible packaging services in the US and has broken ground on its own advanced packaging facilities in Arizona. However, neither of these facilities is expected to come online until 2028 and 2029 respectively. GlobalFoundries, for its part, plans to add additional capacity at its Malta plant to meet demand, but volume production won't begin until the first half of 2028 at the earliest. The partnership is expected to last five years with the option to extend as needed. While CoWoS packaging remains a bottleneck for America's domestic chipmaking ambitions, it's not the only option, just the most popular. Intel Foundry already offers both leading edge process and advanced packaging tech comparable to TSMC's CoWoS offerings. What's more, this tech is compatible with silicon produced at TSMC fabs. Intel routinely employs chips made by TSMC but packaged in its facilities. Doing so does require designing parts for Intel's packaging, but if making chips in America did become a priority, customers wouldn't necessarily have to wait for TSMC and Amkor's facilities to come online. ®

What better way to find out what your system can take than by breaking it? Popularized by Netflix more than a decade ago, chaos testing is the practice of deliberately introducing controlled failures into working systems to test their resilience. Now, a startup that built its business around this practice is using AI to automate more of the testing and troubleshooting process. Gremlin’s Foresight AI, a new add-on to its current “chaos engineering” services, autonomously breaks software infrastructure to uncover hidden bugs before they pop up in production. Chaos engineering can be deliciously dangerous for an engineer: Turn off a random server, load balancer or even an entire region of a cloud deployment, then observe how well the rest of the system tries to keep serving customers. If it does, job well done. If it's borked, more reliability work is afoot. Kolton Andrus, one of the Netflix engineers who did chaos engineering at the streaming service, founded Gremlin to bring fault injection to enterprises, along with business-friendly accoutrements such as scoring, executive reporting, and organizational accountability tools. Adding AI into the mix speeds the process considerably, Andrus told The Register, by automating many of the preparatory and post-testing tasks that used to be done by hand. Failure-as-a-Service AIOps services are nothing new for the site reliability engineer (SRE), but most of these tools diagnose issues only after the system has already crashed. "A lot of AI solutions are, 'Hey, we took a guess. Here you go. Good luck,'” Andrus said. Instead, Gremlin purposefully tips over the system and then offers expert advice on how to prevent that from happening again. To use Gremlin, the user installs agents on their system that can inject latency, kill pods, or max out memory. A Gremlin cloud service then observes the subsequent telemetry alerts for telltale signs of systemic weakness. Gremlin’s secret sauce is its Failure Atlas, a repository of millions of chaos engineering experiments that the company conducted over the past decade on “tens of thousands of systems,” Andrus said. Gremlin developed a harness that yokes the LLM to the Failure Atlas, allowing it to generate more technically grounded answers about what went wrong, rather than simply drawing on its own collection of random information found on the Internet, the company asserts. Gremlin uses a variety of closed and open LLMs for the job, depending on which one is working best at the time. The Atlas keeps the LLM on point, minimizing hallucinations, Andrus said. Not creating problems, just pointing them out Gremlin’s existing guardrails are built around the enterprise’s own access permissions and other security precautions to keep the “blast radius” (company’s words) to a minimum (if the blast isn’t contained at all, well, that’s an access control issue right there). Once a system failure is induced, Foresight AI determines the root cause, generates code or the configuration change it thinks will fix the issue, and then can either apply the solution, or prepare a report for an SRE to read. It’s basically an agentic loop of test-and-replace – keeping humans in the loop at critical junctures – that runs until the failure is no longer induced. Gremlin’s current user base tends to be larger compute-heavy enterprises, many specializing in financial services, retail and enterprise SaaS, Andrus said. They use Gremlin to test disaster recovery plans, ensure correct operation under less-than-ideal conditions, and make sure Kubernetes scales correctly. Such complex distributed systems are particularly difficult to diagnose when they go tits up. Difficult to catch with routine integration and unit testing, some of these bugs can be flushed out of hiding by injecting faults such as network latency, memory exhaustion, or other conditions that expose weaknesses in distributed systems. The corporate rush to use AI to both write applications and deploy infrastructure brings its own set of complications, explained Andrus in an interview with The Register. AI works at such a speed that it is impossible for human reviewers to keep up, and AI can make dumb mistakes anywhere. Any service that promises to break systems for the greater good will require sign-off at the highest corporate levels, said noted Intellyx analyst Jason English, in a note about the technology. For it to work effectively, “we need to change our mindset about risk,” he wrote.®

The credential-hijacking Shai-Hulud worm has struck again, this time burrowing its way into a popular AI agent platform SDK. Multiple security researchers reported Thursday that they had detected Shai-Hulud infection in a recent release of the npm package for version 0.5.144 of Tensorlake’s SDK. That package has somewhere in the neighborhood of 12,000 downloads per week, while its GitHub repository has more than a thousand stars, suggesting it’s quite popular and that the infection could pose a serious risk to anyone who installed the malicious version. Analysis of the malicious release suggests it shares code and techniques with the Shai-Hulud variant dubbed ChainDrop by researchers, which was used in August to compromise npm dependencies including keyv and flat-cache. Like other variants of Shai-Hulud, the worm is designed to steal credentials and self-propagate. This particular version, according to supply chain security firm SafeDep, is designed to steal everything from crypto wallets to browser passwords, GitHub Actions secrets, cloud credentials, service-account tokens, and whatever else it can get its hands on. It exfiltrates that data and keeps an open line to its C2 infrastructure to await further instructions. To make matters worse, this Shai-Hulud variant monitors certain stolen GitHub tokens and, if one is revoked, can trigger the deletion of the infected user's home directory under specific conditions, making removal tricky. Socket recommends rebuilding compromised systems from a trusted source before restoring access to secrets, while researchers warn that the malicious token monitor should be disabled before revoking affected credentials. Tensorlake, for those unfamiliar, is a cloud-native platform for running isolated AI agents and untrusted AI-authored code. The infected npm SDK is used to create and manage Tensorlake environments. Socket warns that the malicious SDK's installation script can execute on the developer's machine or build server, outside Tensorlake's sandbox protections, potentially compromising the host before any AI-generated code is run. “Teams may isolate an agent’s generated code while installing its SDK on a developer workstation, application server, or build runner with access to deployment credentials and other secrets,” Socket noted. “Code executed during that installation inherits the permissions of the installing process.” That may sound bad, but it’s worth noting the malicious version wasn’t up for long - according to security firm Socket, the infected version was published to npm earlier this morning, UTC, and was flagged by its engine 11 minutes after publication. Npm removed the version, and Tensorlake has pulled the package as well, updating the version to 0.5.145. Best check to be sure you haven't installed the malicious version if you're a Tensorlake user. ®

Cryptocurrency transactions lend credibility to parts of a purported leak from Russian extortion crew Silent Ransom Group (SRG), according to blockchain researchers. Chainalysis said some wallet addresses in the material match its existing intelligence, although it could not authenticate the entire collection. It posted: "While we cannot speak to the totality of claims documented in the leak, we can confirm that certain leaked SRG addresses sit downstream of millions of dollars in ransomware payments that SRG has extorted from victims." Despite its name, SRG is known for stealing data and demanding payment rather than deploying ransomware. Its methods include callback phishing and physical intrusions. The FBI warned in May that people posing as IT support staff were entering law offices and copying files onto USB drives. Chainalysis said the transaction histories of two SRG wallet addresses detailed hundreds of thousands of dollars' worth of funds sent from the wallet of a known SRG member. The company believes the funds it observed flowing through the leaked addresses came from a large extortion payment made by a victim in mid-2026, and were used to pay for SRG's expenses, such as members' wages and IT infrastructure. "Our confidence stems from the leaked addresses' transaction history," said Chainalysis. "Several payment addresses contained within the leaks source their funds from confirmed SRG ransom payments. "For example, one of the Silent members' wallets specified in the leaked chats is funded entirely from a $10 million+ victim payment that SRG collected in mid-2026, and which we were tracking before the leak." Blockchain analytics company Crystal Intelligence separately examined the leaked chats and traced payments to wallets identified in them. The chats claim SRG received about $207 million from 27 firms between April and September 2026. Crystal could not verify that total, although it traced funds to an upstream collection wallet that received about 2,675 Bitcoin over its lifetime. Crystal identified a range of ways SRG spent this money. Sometimes the funds would be sent directly to affiliates, but the group had ways of swapping the crypto for cash they could actually use. According to Crystal's analysis, the group used instant exchangers, services that swap crypto for cash sent to Russian bank cards, and for larger deposits, a Moscow-based broker named "Zhenya" who provided the group with physical cash in exchange for crypto. The leaked chats suggest members suspected Zhenya of skimming money through the exchange rate, Crystal said. Crystal said the chats identified a Bitcoin-to-Zelle service advertised on Telegram as SAFU Exchange. Its database associates the same handle with a Georgia-based exchange it describes as unlicensed and sanctions-flagged. The riskiest of its payment methods was reserved for smaller sums paid directly to the likes of "field agents and document forgers," who received their payments directly into their exchange wallets. Crystal said these were regulated exchanges that verify customers' identities. "This is the network's weakest point, and the most direct route for law enforcement to identify the people behind the handles," it said. According to Crystal, the chats included advice to buy property in person and discussions of purchasing new-build homes and sports cars. They said to use mortgages for the homes and, if any banks asked questions about the source of the money, to tell them it was an inheritance or a gift, or to produce a fake loan agreement. The material appeared online this week under the title "The Luna Moth Files," a reference to another name researchers use for SRG. SRG has been active since around 2022 and has maintained a consistent focus on law firms, although it has branched out to other types of organizations too. The group's usual method is callback phishing. Presenting as IT support staff, they convince marks to return their calls and grant remote access to their desktop sessions, running various tools to steal data. The Luna Moth Files website claims the material contains nearly 5,700 internal messages identifying senior members and field agents. Those identities have not been independently verified. These "field agents" are the individuals who were recruited to walk into US law offices to steal data via a thumb drive. Crystal's analysis of the chat logs concluded that these "field agents" were recruited on Russian-language job boards. Job listings promised $300+ per night for "nightclub promoters," although respondents were instead pushed into the physical intrusion line of work. The FBI's May advisory renewed its warning about SRG following fresh reports of attacks that spring. It did not disclose how many incidents involved physical intrusions. ®

More than half of quantum computing startups will be out of business by 2030, Gartner predicts, as an overcrowded market struggles to generate enough commercial revenue. The analyst expects customers to reject technologies that fail to achieve fault-tolerant quantum computing – reliable computation despite errors in the underlying hardware. Despite that forecast, Gartner says enterprises are making initial investments as quantum technologies begin to show early signs of practical business advantage. Broad commercial usefulness remains elusive, but organizations are building expertise rather than waiting for the technology to mature. "CIOs and IT leaders should treat their quantum journey as a multi-year endeavor with the aim of gaining a competitive business advantage over their closest competitors," says Gaurav Gupta, Gartner VP analyst and chief of research for its emerging market dynamics team. Estimates put the number of quantum startups at between 200 and 300, competing in a market Gartner forecasts will generate $1.1 billion in worldwide revenue in 2027. That is a tiny fraction of the quarterly revenue of companies such as Samsung and Nvidia, but still up on the $869.9 million forecast for this year. As The Register reported a couple of years back, developing useful quantum computers requires sustained funding, and while big players like IBM or Google have no shortage of revenue streams, startups may be subsisting on seed funding that will eventually run out if they can't persuade investors that they are onto something. The quantum industry therefore still faces numerous challenges, but continues to make steady progress, not just in the number of quantum bits (qubits) in a system, but in algorithm development, gate speeds, and reliability. And breakthroughs are happening more rapidly than initially expected, at least according to Gartner. "The early quantum advantage era has arrived because noisy intermediate-scale quantum processors now have enough qubits to do limited complex tasks with improved error reduction," claims Gupta. The firm identifies the public sector and the financial industry as the two markets likely to spend the most on quantum computing initiatives. Of these, the public sector is forecast to lead at first, driven by a strong desire for nations to gain technical leadership. Total spend here in 2027 is expected to amount to $245 million, up from $219 million this year. Banking, finance, and insurance firms are expected to overtake the public sector in 2028, spending $289 million against its $280 million. Gartner forecasts that they will remain the largest customer sector through 2030. The US Department of Energy recently kicked off an initiative that will see up to ten participants compete to deliver a fault-tolerant, scientifically relevant quantum computer by 2028, an ambitious goal that is almost certain to fail given the modest amount of funding it is offering. ®

The use of phishing emails as a means of stealing information or implanting damaging malware dates back to the mid-1990s. Although almost as old as the Internet, it remains the instrument of choice for threat actors wanting to pose as trusted organizations or individuals for the purposes of corporate infiltration. In fact its use is on the rise: according to the most recent figures from the Federal Bureau of Investigation (FBI), some 26 percent of all cybercrime complaints filed with them are now phishing-related. The bad news doesn’t stop there. Phishing is mutating in alarming new ways, raising all sorts of difficult questions for defenders and placing email security at a tricky inflection point. AI has for years been a useful tool in the hands of the cybercriminal. But the advent of Generative AI (GenAI) is proving a cyber game changer, transforming phishing from a widespread but easily identifiable nuisance into a lethal precision instrument. Aspiring phishers have historically been hindered by various constraints. Their attempts at chummy authenticity have often been undermined by easy-to-spot mistakes and amateurish formatting – blemishes that traditional email security is good at picking up. Personalised attacks have also been hard to launch at any kind of scale. GenAI sweeps all technical, linguistic, and operational restrictions aside by automating sophistication. Attackers no longer have to choose between individually targeted “spear phishing” which takes much effort to coordinate, and large-scale attacks that lack finesse. Now, with minimal expertise and at low cost, they can hit thousands of victims with minutely tailored phishes at the press of a button. Today’s Large Language Models (LLMs) generate polished and contextual text in any language, backing it with realistic brand graphics and convincing web addresses. At a stroke, Gen AI has democratized this type of crime, putting it in reach of anyone capable of the most elementary research. “Thanks to AI, the historic signals that exposed a phishing email – poor grammar, misspelt words – are now ironed out and replaced with perfect language,” notes Dave Baggett, SVP Cybersecurity with software developer Kaseya. “These emails can be highly targeted with the help of AI. You can create an email that pinpoints, say, a pharma company by using authentic terminology. It’s an incredibly powerful tool for the bad guys.” And let’s not forget that where this new AI-enabled phishing is successful, it is not just a single unwary employee whose data is at risk. Once past the perimeter, a bad actor can get to work burrowing into the cloud and SaaS platforms that now underpin global commerce and communications. All in all, it’s easy to understand how losses from phishing attacks have gone up 274 percent in the last couple of years, according to the FBI. This new email security landscape is an alarming development for corporate IT bosses, and also for MSPs who must be super wary on behalf of customers who don’t have the expertise to pick up attacks at this pitch of sophistication on their own. AI is for the good guys too It’s not all bad news on the email security front line. As threat actors have been perfecting their use of GenAI, defenders have been developing AI-driven counter solutions in parallel. Where old school protection relied on spotting the kind of anomalies that GenAI has now ironed out, newer solutions are geared more towards contextual analysis. They are not so much trying to spot clumsy typos and dodgy attachments as model an attacker’s intent before damage is done. Today defenders can use AI to analyze subtle patterns that are in tune with the layered structure of modern phishing campaigns. The aim is to determine whether a message aligns with expected behavioral patterns. This level of finesse is made necessary by the insidious nature of modern phishing which means that technically clean email messages, in other words ones that are free of malware-loaded attachments or booby-trapped links, can still contain malicious potential. Modern cybercriminals often bypass automated security filters by the simple means of exploiting human psychology and trusted infrastructure. Social engineering, supercharged by GenAI, can be a sharper sword than a malware payload. A phish will often seek to impersonate a boss or a colleague using a plain text email. Attackers can set GenAI to scour LinkedIn, looking for people who have just started new jobs. Green employees are more vulnerable to a phish that looks like an email from a new superior they haven’t got to know yet. Once trust is established, what will typically follow is a demand for, say, an urgent wire transfer, or a casual request for payroll details or passwords. Attackers can develop a relationship over multiple emails to “groom” the victim before striking. Traditional filters see nothing suspicious. Spotting this sort of attack demands a smarter grade of tool. “Since attackers are relying on manipulation of fallible individuals, security must now support them at the moment where wrong decisions can be made,” says Baggett. “Good email security will replace generic warnings with easily digestible context about why a message might be risky and advice about the need for further verification.” Where once an inbox was a passive delivery channel, says Baggett, it is now an active layer of risk mitigation, tooled up to stamp out a phishing attempt before it escalates into account compromise and devastating financial loss. A huge asymmetry The good guys are fighting back. But as Baggett points out, there remains a huge asymmetry between bad actor and defender: “Attackers can use LLMs, basically for free, to create perfect phishing templates,” he says. “On the other hand, if defenders took every inbound email and put it through a frontier LLM model, that’s about 100x too expensive. We just can’t use the same tools at the criminals.” One option for defenders is to use smaller, more specialized models, distilled from larger ones. A compact model could be tuned, for example, to the sentiment or meaning of each phrase in an email, and set to pick up a threatening tone or a favor being asked. Where chunky LLMs just cost too much, another choice might be to run a subset of the overall inbound mail through a big model: “Admins can be given a way to run particular mails through a frontier LLM,” says Baggett. “That model will have been trained to understand certain critical aspects of email security.” Better use could be made of pre-LLM tools such as computer vision, which can help process, analyze and understand visual data like digital images and videos. Other ancillary tools that could enrich the defensive mix include sender analysis to look into the origin and authenticity of an email message, and linked-content inspection that can examine an email in depth without triggering any embedded threats. There are some reasons to hope for a safer future as AI technology evolves. As inference becomes smarter and models get better, it should become possible to run a higher percentage of emails through a large scale LLM, believes Baggett. This, however, may not be achievable for some years. There’s also the possibility of using LLMs not just to spot potentially malicious emails but also to analyze the settings and admin controls of existing security tools. In this way, the knowledge required to use the tools to best effect becomes much less. “Security tools can be complex and feature hundreds of settings,” points out Baggett. “Some require expertise that our customers don’t have. There are a number of things we believe tools should be doing to help defenders, for example reduce alert fatigue.” The criminal fraternity may, of course, also find ways to harness better tools for new purposes. At the moment, most attacks are conceived and designed by humans. But ultimately, this job may be taken over by agentic models able to figure out new tactics for themselves. Machines may soon be able to plan and supervise attacks at scale. Great tools, available now Luckily defenders don’t have to sit back and be at the mercy of future developments. There are some great tools available today that help them even up the odds and cut the complexity from checking email. INKY Smart Insights, for example, uses GenAI to reduce a lengthy manual email investigation to a few seconds of automated triage. For the purposes of hard-pressed administrators it can turn a mess of technical email evidence into a plain-language verdict and enable the creation of cogent reports. Along with Kaseya Intelligence it can support decision-making across the broader security environment. “Smart Insights will identify problems and give a narrative explanation of its reasoning,” enthuses Baggett. “It will do that in a way that’s not overly technical, making the tool valuable to someone who doesn’t have expertise in email security.” The hard-pressed MSP, for example, can leverage Smart Insights to provide key information for their customers in an abbreviated and digestible format. In short, it allows them to enjoy the sophistication of LLMs but without adding hugely to their overheads. The war on GenAI-powered phishing is far from won. But so long as defenders can understand the potential of the smart tools at their disposal, they have a fighting chance of keeping the bad guys at bay and essential channels of communication safe. Find out more at Kaseya’s Cybersecurity Summit – details here Sponsored by Kaseya

"Forget It" was the name of a small computer program never thought to be needed, which will nonetheless go down in history along with its author, Margaret Hamilton, the American software engineer who died on September 30, aged 90. Hamilton led development of the onboard flight software for Apollo's command and lunar modules, helping astronauts reach the Moon in 1969. Among her early jobs after joining the MIT Instrumentation Laboratory in 1965 was to write the software for the Apollo abort procedure. "Everybody said that was never going to happen, [so they said] 'Oh well, good. We'll give this one to Margaret because she's a beginner,'" she told the Computer History Museum. However, an unmanned flight did abort, raising Hamilton's standing within the team. "It went to [run] this program I had written, which I named Forget It. All of a sudden, I became an overnight expert," she said. Hamilton was born in Paoli, Indiana. Moving with her family to Michigan, she graduated from high school in 1954. She studied mathematics at the University of Michigan and Earlham College, where she graduated in 1958 with a minor in philosophy. She planned to pursue graduate studies and an academic career, but decided to go to Boston in 1959 with her husband, who was studying law at Harvard. She got a job at MIT with Edward Lorenz – who went on to pioneer chaos theory – and programmed the LGP-30 and DEC's PDP-1 computers for weather forecasting. She worked on other projects at MIT, including aerospace systems, before hearing of the Apollo missions in 1965. She immediately applied to join the MIT team supporting them. She initially wrote software for uncrewed Apollo missions, later leading development of flight software for the crewed program. The team's software played a crucial role in the first lunar landing as the Apollo 11 module approached the Moon's surface in 1969. During the final minutes of descent, the guidance computer raised several unexpected alarms, leading the astronauts to question whether it was safe to land. The alarms showed "executive overflows," meaning the system could not complete all its tasks. Hamilton had also pressed for safeguards against operator error. Her daughter, four years old at the time, crashed a command-module simulation by activating a pre-launch program during a simulated flight. Hamilton proposed protection against that mistake, but was initially overruled because astronauts were considered unlikely to make it. An astronaut subsequently did so during Apollo 8, prompting changes to the software. For Apollo 11, the software's restart protection allowed essential landing tasks to continue despite the overload. Computer engineer Jack Garman advised guidance officer Steve Bales that the alarms did not require an abort. "We're go on that alarm," Bales confirmed. The crew made it safely to the Moon's surface and later returned to Earth. The overload was traced to unwanted inputs from the rendezvous radar system. While working on the Apollo missions, Hamilton began to call herself a "software engineer" because writing code was not taken as seriously as other engineering disciplines. The term is now standard. She also features in an iconic photograph, standing beside a stack of her team's Apollo software listings almost as tall as she was. Following the Apollo missions, she founded and co-founded a number of technology companies focused on fault-tolerant software. She created the Development Before the Fact (DBTF) methodology to prevent errors early on, along with the Universal Systems Language (USL) for systems design. In 2016, President Barack Obama awarded her the Presidential Medal of Freedom. Computing pioneer Grace Hopper received the same honor posthumously. The White House described Hamilton as a mathematician and computer scientist who contributed to concepts of asynchronous software, priority scheduling and priority displays, and human-in-the-loop decision capability, which set the foundation for modern, ultra-reliable software design and engineering. Hamilton's legacy is a reminder to design for the mistakes users supposedly won't make – and the failures no one expects. She is survived by her daughter, a son-in-law, two grandsons, four great-grandchildren, and three siblings, according to MIT. ®

ArtCraft has added alternatives to Microsoft Word, Excel, and PowerPoint to its growing collection of Rust applications, which it describes as "clean-room reimplementations" of proprietary software. The projects lean heavily on Claude for development and remain unfinished, but our first look at its Word alternative, WordCraft, suggests they are worth watching. The company's GitHub repositories also include projects modeled on Photoshop, Illustrator, and Premiere Pro. We fired up WordCraft on an Arm-based Surface and it worked pretty well. WordCraft resembles Microsoft's application, although its buttons, icons, and other imagery differ. "The spelling dictionary and hyphenation come from Grady Ward's public-domain Moby Hyphenator II word list," ArtCraft says, and the application is dual-licensed under MIT or Apache 2.0. The company is at pains to point out that WordCraft is at the pre-alpha stage and estimates that it is at 62 percent of real feature parity, although "an alpha for everyday writing is close." We would keep it away from production workloads for now. Its roadmap says testing DOCX compatibility against a collection of real-world files has not begun, and charts, SmartArt, and embedded OLE objects are dropped. The Document Foundation, which develops LibreOffice, declined to comment. Microsoft acknowledged our inquiry but offered no further response. Adobe also declined to comment on the projects targeting its creative applications. Neil Brown of decoded.legal told The Register that the legal position would depend on the jurisdiction, but suggested that "those doing the 'reimplementations' may well argue that, since they have implemented the underlying functionality independently (i.e. without access to the source code of the software that they are reimplementing), their overall reimplementation is non-infringing." That said, "the legal situation is always going to be interesting, and quite possibly litigated, when the party most likely to feel aggrieved has deep pockets and lots of lawyers, and probably quite a lot to lose." ArtCraft says WordCraft was built from public specifications and observation, using original or openly licensed assets. Its ambitions now extend from creative software to office applications, but whether the projects can deliver reliable replacements remains to be seen. In the years to come, it would be richly ironic if AI programming tools took a bite out of the revenues of the company that gave us Copilot in GitHub. ®

CrowdStrike researchers investigating attacks on South Korean financial institutions found exposed AI session logs containing operational details and a resume-writing request that may identify the attacker. In a report published Wednesday, analyst Ashley Campion said the prompt named "YY," listed a Chinese university and a location in Guangdong, and gave conflicting age information: 26, but initially a birth date in September 2007. CrowdStrike considers the details likely to belong to the attacker but says it cannot definitively establish that connection. The attacks affected at least five lenders, including Shinhan Bank, KB Kookmin Bank, Hana Bank, Yegaram Savings Bank, and BNK Busan Bank, according to The Korea Times. In one case, the attackers allegedly breached a loan progress inquiry service and in another, they gained access to a mobile work-support system. The researchers found references to YY in AI sessions associated with activity involving ARTEX, a recently released open source penetration-testing tool developed in China and used in the attacks alongside Claude Code. Researchers examined an exposed directory on a server associated with the attacks. A Chinese-language instruction file led them to another server in Hong Kong, where they found session histories, configuration files, and AI memory files documenting the targeting. The resume prompt included a Telegram username that also appeared in activity targeting a possible Chinese payment platform and in Claude Code sessions seeking vulnerabilities in a Telegram-based NFT gift marketplace, CrowdStrike said. "The use of agentic AI tooling alongside traditional offensive capabilities highlights the continued evolution observed by CrowdStrike in adversarial tradecraft," Campion said. "This activity demonstrates how AI tooling can enable a financially motivated threat actor to conduct multiple intrusions within a short time span. CrowdStrike Intelligence assesses that adversaries will likely continue to experiment with implementing AI tooling in their operations to enhance their operational tempo and capabilities." Police are investigating whether an individual or an organized group carried out the attacks. Shinhan Bank reported that about 25,000 customers were affected, while KB Kookmin and Hana reported 119 and 89 respectively. Lawmakers have approved plans to summon the heads of five major commercial banks to an October 19 parliamentary audit to answer questions about cybersecurity lapses. ®

IBM has sold “hundreds” of its cloudy VMware customers to Broadcom partner 11:11 Systems. Under Broadcom’s ownership, VMware slashed the number of cloudy partners it keeps on the books and retained only partners that go all-in on the Cloud Foundation (VCF) private cloud stack. Partners that Broadcom decided not to work with lost the right to resell its software as of October 2025 and may not run VMware-powered clouds after March 2027. IBM was once a top VMware partner, but last year stopped selling new cloudy VMware services, suggesting it wasn’t invited to continue as a Broadcom partner. Sources tell The Register that soon-to-be-former VMware partners often have a contractual obligation to provide a Virtzilla-powered cloud beyond the March 2027 cutoff, and that their customers want whatever happens next to be non-disruptive. One option for partners leaving the VMware fold is to turn to a wholesaler like 11:11, which runs its own VMware cloud. Another option for departing partners is selling contracts for ongoing VMware cloud services to one of the remaining Broadcom partners. 11:11 Systems told The Register IBM sold contracts for “hundreds” of customers who use VCF in the Big Blue cloud. 11:11 CEO Brett Diamond said the contracts are for customers running versions 7 and 8 of VCF and who are therefore “ripe for migration.” IBM did not respond to our questions about the transaction, which included a request for information regarding some customers that we understand are not making the move to 11:11. Diamond said 11:11 has 40 to 50 points of presence for its VMware cloud, and a migration methodology it will apply to the customers it bought from Big Blue. He’s confident his newly acquired customers will want to stick around. “The customers we manage are very large multinationals and Fortune 5000s who want to stay on VMware,” he said. “No other hypervisor can handle the mid-to-large enterprise.” The CEO hopes to shift his new customers to VCF 9, a necessity given VMware ends support for VCF 8 next year. 11:11 also uses its VMware cloud to sell other services, by delivering it through a custom console that integrates around 50 other tools spanning backup, storage, and security. The IBM buy is 11:11’s eight acquisition of a VMware service provider. It’s not hard to guess why Broadcom binned Big Blue, given that it sells Red Hat OpenShift which is a competitor for VMware – and a stronger one after recent improvements to its VM-hosting capabilities. Other VMware partners feel Broadcom did not have clear-cut reasons for excluding them from its partner program. Broadcom argues that services outfits that don’t dedicate their practices to VCF won’t meet customer needs, and it is therefore building a better channel composed of expert partners. Lobby group CISPE (Cloud Infrastructure Services Providers in Europe) argues Broadcom’s actions mean it has improperly concentrated the market for VMware cloud providers and is pursuing an antitrust case in EU courts. ®

Ten of the world's biggest AI developers have agreed to tighten their handling of personal data following scrutiny by Britain's privacy watchdog, which is now turning its attention to autonomous AI agents that don't always play by the rules. The recently rebranded Information Commission's Office (ICO) said Thursday that Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI, and Stability AI have either made changes or committed to doing so after the regulator examined their compliance with UK data protection law. These include clearer explanations of how personal information is used to train AI models, better ways for people to exercise their data rights, and tougher assessments of developers' safeguards. The commitments follow a supervisory program launched in 2025 that initially covered 11 developers. That number dropped to ten after the ICO paused its engagement with Elon Musk's xAI to pursue a separate formal investigation into its Grok chatbot. The watchdog isn't declaring victory just yet. It's monitoring whether developers deliver on their promises and admits that current AI training practices still pose problems under UK data protection law. Developers still have some explaining to do over personal data buried in their models, particularly sensitive information, and how people are supposed to get their details removed once an AI has been trained on them. There's also the risk of personal information being extracted from models, including data developers never intended them to retain. The ICO acknowledged that some of these issues will require cooperation between industry, regulators, and government. Getting companies to promise changes is one thing. Making today's AI models comply with the law is another. "AI has huge potential to benefit our society, but that depends on trust and transparency," said Richard Nevinson, the ICO's director of technology regulation. "Our engagement with some of the biggest developers has secured real commitments that will help people better understand and control how their data is used, even in a fast-moving and complex area." Meanwhile, the regulator is turning its attention to AI agents, which can browse websites, use tools, and carry out tasks with limited human supervision. And some aren't sticking to the rules. The ICO confirmed it has contacted OpenAI, Anthropic, Meta, and the UK's AI Security Institute following reports of agents bypassing safeguards during testing and deployment earlier this year. "These recent reports show both how fast these systems are advancing, and the risks they pose if the guardrails aren't fit for purpose," Nevinson said. "Our message is clear: the fact AI agents act with autonomy is not an excuse for poor compliance." The regulator has also launched a six-week call for evidence on agentic AI, covering security, transparency, accountability, and the lawful use of personal data. Responses are due by November 20 and will inform future guidance and the ICO's forthcoming statutory code of practice on AI and automated decision-making. The watchdog is separately examining how consumer chatbots and AI companions use personal information as they become increasingly personalized. For now, the ICO has secured promises from some of the industry's biggest names, although whether they deliver remains to be seen. And with AI agents increasingly capable of acting on their own, the watchdog may have its work cut out keeping them in line. ®

Samsung Electronics expects third-quarter revenue to more than double and operating profit to reach nearly nine times last year's level as the AI infrastructure boom fuels demand for memory. The South Korean company is among the major memory manufacturers benefiting from demand for chips used in datacenter servers and GPUs. Although Samsung will not report its full earnings until the end of this month, its guidance puts sales at between ₩194 trillion and ₩196 trillion (about $145 billion) for the quarter ended September 30, 2026. This is more than double the ₩86.1 trillion ($63 billion) for the same period last year. The company expects operating profit of between ₩107.3 trillion and ₩107.5 trillion (about $80 billion), compared with just ₩12.2 trillion ($9 billion) for Q3 2025, roughly 8.8 times last year's figure. Samsung's brief guidance note does not break down performance by business division. Demand for memory, including high-bandwidth memory (HBM) used alongside AI processors, is likely to be a major contributor. The demand for DRAM and NAND for AI applications is driving profits across the semiconductor industry. Analyst Gartner forecast that the sector as a whole will bring in $1.6 trillion during 2026, almost double the revenue seen in 2025. Rival chipmaker Micron posted similar gains last week, and the other Korean memory giant, SK hynix, will likely have the same story to tell when it discloses its own quarterly figures later this month. The boom is proving costly for buyers of everyday electronics. Memory manufacturers' shift towards more lucrative AI products is squeezing supplies for PCs, smartphones, and other devices, driving up component prices. As The Register previously reported, rising component costs are pricing budget PCs out of the market. US sales in that bracket fell 18.7 percent year-on-year in the first quarter, while average PC selling prices were forecast to rise by up to 12 percent by December. Smartphone makers face similar pressures, with budget devices becoming increasingly difficult to produce profitably. Samsung warned earlier this year that it expected the memory supply crunch to deepen in 2027 and potentially persist through 2028, leaving buyers with little to smile about. ®

Britain and Germany have announced a partnership to counter cyberattacks and sabotage, particularly from Russia, as Prime Minister Andy Burnham heads to Berlin for talks. Under the arrangement announced Thursday, the countries will share information and coordinate efforts to monitor, deter, and disrupt hostile activity, including threats to critical infrastructure. The announcement comes ahead of Burnham's first face-to-face meeting with German Chancellor Friedrich Merz, where security won't be the only item on the agenda. The PM is also expected to raise Britain's relationship with the EU, including the possibility of rejoining the bloc. Burnham has floated several options for undoing parts of Brexit, from rejoining the customs union or single market to returning to the EU outright. According to the Financial Times, he wants Germany's backing for closer economic ties before a UK-EU summit on November 20. On the security front, both governments worry about Russia's appetite for cyberattacks, sabotage, and other operations that fall short of outright military confrontation. "Attacks on our infrastructure, our businesses and our democracies are not a distant threat," Burnham said ahead of the meeting. "They reach into people's homes and workplaces, and they are designed to make us weaker and more divided." He added: "Britain and Germany are joining forces to find, expose and disrupt this activity." The partnership builds on the Kensington Treaty, signed in July 2025 and ratified on Thursday, which already committed the two countries to working more closely on defense, cybersecurity, technology, and economic security. The announcement offers little operational detail. It names no agencies, specifies no additional funding, and does not explain how the countries will disrupt hostile operations. The deal lands amid increasingly stark warnings from European intelligence agencies about Moscow's activities. Germany's foreign intelligence chief, Martin Jaeger, warned this week that the confrontation with Russia had entered a more dangerous phase, while German authorities have blamed Moscow for an attempted drone attack at Leipzig airport in August. Britain has been dusting off its own tools for dealing with hostile states. The government has fast-tracked legislation giving the Home Secretary new powers to designate state-backed organizations and proxy groups, while Burnham has asked security chiefs to establish a National Centre for Information Defence to tackle foreign disinformation campaigns. ®

Microsoft's latest Windows Search preview promises faster results and less clutter, while adding a new sideline: carrying out commands such as muting audio or minimizing windows. Following July's improvements, Microsoft began rolling out the new preview yesterday. The Actions feature lets users change settings directly from Windows Search. Typing "mute" into the search box and selecting the resulting action silences the speakers, with a "Done!" message confirming the change. The same applies to other actions. We tried "minimise all my windows" (note the non-US spelling) and Windows performed the Show Desktop action. Search is now more results-focused, rather than showing a large preview pane, and there is better handling for typos and synonyms (although there is clearly more work to do: "turn on" and "switch on" confused the preview in our testing). Inline previews and richer file details will follow later this year, helping users distinguish similar results without restoring the separate preview pane. The Windows Search experience is built on WinUI 3 and Microsoft claims "substantial improvements in performance and memory usage." However, this is still an English-only preview, gradually rolling out to Windows Insiders in the Experimental channel. Eligible users can enable it sooner through the Windows Insider Program's Feature flags settings. Faster, less cluttered search results are welcome in a tool that has long been a Windows punchline. Whether a box marked Search should also double as a command launcher is another question. ®

Microsoft launched some new devices last night, but the big news appears to be a rebrand of the Surface range as "Surace." The typo (we assume) appeared on a slide about the number of ports on Microsoft's expensive hardware. Pavan Davuluri – he who promised that Microsoft would deal with Windows' quality woes – was showing off some of the features of the Surface Laptop Ultra when the slide gave the in-person audience and online viewers an unintended demonstration of Microsoft's quality control. We asked Microsoft whether the spelling was intentional, but the company has not responded. Microsoft has had worse presentation mishaps. Windows 98 famously blue-screened during a live USB demonstration. Misspelling the product name requires rather less technical effort, though. We look forward to seeing the results of Davuluri's efforts to fix Windows' woes once he's fixed the typo in the product name on the slide behind him. ®