❌

Normal view

Received — 18 November 2026 ⏭ The Register - Security

California tells Waymo to take the next exit after freeway mishaps

18 November 2026 at 17:45
Waymo has suspended driverless freeway rides across the US after California regulators intervened over robotaxis entering construction zones and trying to bypass highway patrol vehicles. California's DMV banned Waymo from offering driverless trips on its freeways on Wednesday, November 11. The same day, Waymo pulled the handbrake on its freeway services countrywide. The move affects all areas in which Waymo operates on freeways, including San Francisco, Los Angeles, Miami, Phoenix, and Austin. The Alphabet-owned company has not indicated how long the suspension will last. Waymo says freeway testing will continue with trained specialists behind the wheel. Jonathan Groveman, spokesperson for the California DMV, told the Associated Press that the organization imposed the ban following two incidents involving Waymo's robotaxis that posed an "imminent hazard to public safety." On November 5, three of the company's driverless cars entered an active construction zone on US 101 near Burlingame, south of San Francisco. The second incident occurred the following day, and in footage shared widely online, a Waymo vehicle can be seen trying to drive around highway patrol vehicles conducting a traffic break – a maneuver used to slow or stop traffic. "The Department determined that Waymo's driverless freeway operations pose an imminent hazard to public safety due to recurring incursions by its vehicles into active construction zones where construction workers, vehicles and equipment are present," Groveman said in the statement. Groveman said the DMV had already imposed restrictions and required corrective action following similar incidents on May 18. Waymo subsequently recalled nearly 4,000 cars after some were caught ignoring road closure signs and cones designating construction zones in Phoenix and San Francisco. A separate recall covered software affecting nearly 3,800 robotaxis over problems including slowing on high-speed roadways and driving into floodwater. A Waymo spokesperson said that despite California's temporary ban, its robotaxis continue to operate on city streets across the US. "Safety is at the foundation of everything we do," they said. "Following recent freeway construction zone encounters on US 101 in California that did not involve any injuries or collisions, the California DMV has implemented an operating restriction on freeway operations in the state. "In alignment with our safety-first principles, we have also temporarily restricted freeway operations across our other operating markets to evaluate our performance. Our ride-hailing service continues to operate safely on surface streets, serving over half a million trips each week. "We are working with the California DMV and other state officials diligently and responsibly to resume freeway operations." Other nuisances and hazards caused by errant Waymos include clogging traffic in thick fog, driving into telephone poles, and killing cats and dogs. ®

Microsoft's official Python agent framework hijacked to distribute malware

18 November 2026 at 17:23
Microsoft’s official agent framework package hosted on PyPI has been hijacked and now installs info-stealing malware, multiple security research firms have discovered. CEO Satya Nadella may be on stage trying to wow investors and customers at his company’s annual Ignite conference, but somewhere in the bowels of a building in Redmond his open source team is likely panicking at what appears to be unauthorized access to an administrator account on Microsoft’s Github organization. The problem package is Microsoft’s agent-framework, versions 1.26.1 and 1.26.2, researchers reported late yesterday and this morning. According to OX Security, version 1.26.1 was reverted to a safe version, but the attacker later uploaded 1.26.2 through Microsoft’s official PyPI publisher account, suggesting that, at least at the time of their report, the credentials hadn’t been locked down. Safe to say, but if you find either version in your environment assume you have some account security work to do. Microsoft agent-framework for Python boasts nearly 200,000 downloads a month and is used to build and deploy AI agents. As of those latest releases, it’s also being used to exfiltrate environment variables, credential files, and other secrets for Azure, Git, Docker, Hugging Face, and other sensitive developer services. To make matters worse, the package never needs to be imported to do its dirty work – all it takes is a Python interpreter with the package installed to start up – the infected package still pulls the genuine 1.26.0 version of Microsoft’s agent-framework in order to keep itself running and looking legitimate, while in the background it gathers up credentials, encrypts them as a base64-encoded chunk of JSON, and sends them off to a webhook.site address hard coded into the malware. It appears from researcher accounts that this is made to act fast, as the malware isn’t persistent and only runs as long as the malicious version of the agent-framework remains in a site-packages directory. What it does do, however, is not just look for credentials in the environment: It actually starts calling for new ones and collecting them, too. In particular, the malware requests fresh identity tokens from GitHub Actions, Azure DevOps, and Azure itself, which are exfiltrated to the same webhook site. “That combination of passive theft and active token requests makes it especially dangerous in cloud build environments, because even short-lived tokens are useful when they're exfiltrated right away,” said researchers at OpenSourceMalware. The agent-framework GitHub repo is offline as of writing. It’s been pointed out online that Microsoft’s repo for ONNX Runtime, an AI accelerator, is also offline. Other Microsoft repos remain available, and neither Microsoft nor security researchers have linked the ONNX repo’s pull to the malware campaign. “We have made some repositories private while we investigate malicious content,” a Microsoft spokesperson told The Register. “Our priority is to protect customers and the broader ecosystem. We are working to review and will take further action as needed.”® Updated at 1805 with comment from Microsoft.

Cyber insurers now want proof you can actually recover

18 November 2026 at 16:00
Cyber insurance pricing has cooled. Marsh’s Global Insurance Market Index shows global cyber rates falling for a 13th consecutive quarter, with US rates recently stabilizing after the ransomware-driven spikes of 2020 and 2021. Part of the reason rates are lowering is because carriers are imposing stricter requirements that determine if you will qualify at all. A major global carrier recently stopped writing law firms altogether because of claims volume, and brokers expect healthcare and managed service providers (MSPs) to be next. The newest requirement is the one many MSPs are least ready for: proof that you can actually recover. This is what insurers really want to see, and meeting the bar is harder than it sounds. Why insurers stopped taking backups on faith In a recent webinar with cyber insurance provider Beltex, the familiar list of requirements came up first: Multifactor authentication (MFA) on external access and privileged accounts Endpoint detection and response (EDR) on every endpoint A staffed 24/7 team where managed detection and response (MDR) is required, and timely patching. Backup is now joining that list, and for good reason. For years it was assumed that having backups and a green job status was good enough. Claims have shown otherwise. Now, attackers go after backup infrastructure to make their ransomware more effective, and an untested restore can re-infect the production network. Beltex described an MSP that restored a client directly from backup without involving the insurer. The attackers were still in the environment, and the ransom reportedly doubled. So, carriers are asking for proof. Some want evidence of a successful restore within the last six months, and some insurance applications must be signed by a C-level executive rather than the IT director to put ownership behind it. Proof of speed matters as much as proof of recovery Being able to recover is only half the question, because time is where claim costs balloon. Insurers, it turns out, don’t object to paying claims. They object to paying big ones. Beltex’s favorite example involved an 85-person firm hit through a zero-day firewall flaw. Their MDR solution contained the intrusion in about 20 minutes, forensics confirmed two servers were touched, and clean backups had the firm fully functional within 36 hours. The claim came in at around $85,000. Compare that with downtime and business interruption: a firm billing $10,000 a day that is offline for a week loses $70,000 in income before recovery costs and claim submission. Yet large servers can take four to 24 hours to restore. That is why underwriters care how fast you can come back and how successfully. Where you recover is the other half Speed raises a second problem: where do you restore to? Breach counsel treats the first 72 hours as the most critical window, and incident response (IR) teams need to isolate and investigate infected systems during it. Restoring into production too early risks reintroducing the threat and disturbing the evidence investigators need. Waiting for a safe place has the clock running. Recovery that is fast but not safe doesn’t satisfy anyone. Meeting all of this is harder than it looks Put the pieces together and the bar is high: Backups you have actually tested Recovery times you can state and meet And a clean, separate place to recover to Underwriters want to know whether you exercised reasonable care, so a row of green checkmarks won’t do. Run a test restore. Confirm you’re protecting everything that matters, including the server someone labeled “not in production.” Have the business owner, not just IT, confirm the right systems came back within agreed recovery time objectives (RTO) and recovery point objectives (RPO), and keep the logs, screenshots and sign-off in your ticketing system. It doesn’t need to resemble a government audit; Beltex noted that this alone puts you ahead of most organizations. Where cloud disaster recovery fits Cloud-native disaster recovery changes the equation. Cove’s DRaaS solution lets teams stage a pre-recovered VM in the N-able cloud that can be failed over to in about 15 minutes, off the production network. That answers the speed question and the location question together. Forensic investigators can validate and hunt in the sandbox while the business keeps moving, with rollback to earlier points in time. Once everyone agrees the recovery point is clean, workloads can run from the cloud over VPN, and failback uses recovery agents to bring changes home. Where to start Insurance is a financial failsafe, not a substitute for a bad data protection posture. Start by listing what the business couldn’t lose for 72 hours, set recovery targets, test them and document the results. Check the policy itself: a standalone cyber policy rather than a small add-on to a general one, with a limit of at least $1 million, and $2 million as the business scales. And if an incident does hit, disconnect affected machines from the network rather than powering them off, so forensic data survives. N-able Cove Data Protection, recognized as a Champion in the Infotech 2026 Enterprise Backup Data Quadrant, provides cloud-native data protection for organizations. Learn more at n-able.com/products/cove-data-protection. Contributed by N-able.

AI-powered cyber threats are changing the rules of enterprise recovery

18 November 2026 at 15:00
Cybersecurity has long been a contest between attackers finding new ways in and defenders trying to keep them out. Artificial intelligence is making that contest more complicated, giving cybercriminals new ways to develop and adapt their attacks while putting additional pressure on conventional cyber defences. For enterprise IT leaders, the challenge extends beyond preventing a breach. What happens when an attacker gets through? Can compromised data be identified, clean copies recovered, and essential systems brought back online without prolonged disruption? These questions are becoming increasingly important as organizations confront both traditional threats and emerging AI-powered cyberattacks. A successful defense strategy must, therefore, account for what happens after an incident, not just the measures designed to prevent one. Making recovery part of the defence This is where cyber-resilient enterprise storage has an important role to play. Rather than relying solely on traditional backup processes, organizations need ways to protect critical data, detect potential compromise, establish which copies can safely be recovered, and restore those copies as quickly as possible. Infinidat addresses this challenge through its InfiniBox and InfiniBox SSA enterprise storage platforms, which incorporate the company’s InfiniSafe cyber resilience and recovery software suite. InfiniSafe combines immutable snapshots, which cannot be altered or deleted during their protected retention period, with logical air-gapping to isolate the management plane from the data plane. Fenced forensic environments provide a separate space for validating data before recovery. The technology also supports automated cyber protection, allowing security monitoring data center wide cyber security software applications or a Security Operations Center (SOC) to automatically talk to InfiniBoxes and trigger immutable snapshots when suspicious activity is detected. This helps reduce the window in which critical data might be exposed. For organizations concerned about recovering compromised information, optional InfiniSafe Cyber Detection uses AI-based analysis to help identify cyber affected data and locate the last known good copy of the data for recovery. Speed is another consideration. Infinidat guarantees recovery of immutable snapshots in one minute or less for InfiniBox and InfiniBox SSA, subject to the guarantee’s terms and conditions. Ultimately, cyber resilience isn’t about assuming every cyberattack can be stopped. It’s about ensuring that when defences are breached, organizations have the tools and processes to recover trusted data and limit the operational consequences. For enterprise IT teams, that makes recovery readiness an essential part of the cybersecurity conversation. Sponsored by Infinidat

UK MPs want AI watchdog with power to pull the plug on dangerous models

18 November 2026 at 09:30
UK MPs want a watchdog with the power to pull the plug on dangerous AI models, rather than leave the likes of OpenAI and Anthropic to decide when they've gone too far. In a report shared with The Register ahead of publication on Wednesday, a parliamentary committee is calling for an AI Safety and Security Bill that would give a watchdog powers to intervene while models are still being developed. It could halt high-risk training and testing, block deployments, or suspend AI services if the risks can't be contained. LLMs are trained on billions of learned parameters and so developers can't always predict what a model will do in every situation. The recommendations come from Parliament's Business, Innovation, Science and Trade Committee, which questioned researchers, former AI employees, and representatives from OpenAI, Google DeepMind, Anthropic, and Meta. All four companies assured MPs they wouldn't release models they considered unsafe. The trouble is, they don't use the same yardstick for safety, and UK law doesn't generally require them to get an independent opinion before releasing anything. "We cannot give today’s AI leaders unfettered freedom to behave like modern-day Dr Frankensteins, creating dangers beyond anyone’s control," said committee chair Liam Byrne MP. Just as with aviation, pharma or nuclear safety, UK government want to rely on manufacturers' assurance alone, and want regulation. "Our hearings exposed gaps in independent scrutiny, incident reporting and powers to stop unsafe systems. Once some models are released, they cannot reliably be recalled," added Byrne. "Protection must begin while systems are being built and tested." The committee's report, Guardrails for Growth: Making AI Safe, identifies ten holes in Britain's approach to AI safety, including the lack of mandatory incident reporting, limited access for independent testers, and inadequate protections for whistleblowers. The UK already has an AI Security Institute (AISI), which has made a name for itself testing some of the world's most powerful models, but its work depends on voluntary cooperation, and it has no authority to stop a company releasing a model that raises safety concerns. MPs want AISI to continue its testing work while a separate regulator takes responsibility for enforcement. Developers of designated high-risk systems would have to provide evidence supporting their safety claims and give independent researchers access to examine them. Significant changes to a model's capabilities or safeguards would trigger further scrutiny. Former AI employees also told MPs that raising safety concerns can come at a price. Restrictive contracts, commercial pressures, and financial disincentives can all discourage staff from speaking out, the committee heard. MPs want stronger legal protections for whistleblowers and compulsory reporting of AI incidents and near misses. There's a fair bit of money riding on this, too. Using OECD productivity estimates, the committee calculates that AI could leave Britain's economy up to £400 billion larger over the next decade. It also sees an opportunity for the UK to make money from independently testing AI models, building on the expertise already established at AISI. MPs argue that stronger safeguards could encourage businesses, particularly smaller firms, to adopt the technology. "AI firms must be required to substantiate their safety claims, independent experts must be able to scrutinize them, and a regulator must have the power to intervene when risks cannot be controlled," Byrne said. "Ministers, accountable to Parliament, must decide what risks the public should be asked to bear." MPs want the new laws on the books in the next parliamentary session, and say Britain should use its turn at the G20 helm in 2027 to get other countries on board with tougher rules for AI testing, incident reporting, and risky model releases. Whether ministers are prepared to hand a watchdog the power to stop AI development is another matter. The committee's recommendations aren't binding, and the government has yet to commit to the proposed legislation. ®

Oracle reveals fix-this-ASAP-grade Java4Shell bug that allows attacks on many third-party apps

18 November 2026 at 04:33
The United States Cybersecurity and Infrastructure Security Agency (CISA) has urged Java users to immediately patch a high-severity flaw in the software, which attackers can use to crash the Java virtual machine or achieve unauthenticated remote code execution (RCE). Oracle revealed the flaw in its monthly patch dump for November, which offers 655 patches, 62 of them rated critical, affecting 62 products. Several of the flaws are rated more dangerous than the Java problem, CVE-2026-114938, which Big Red didn’t single out as worthy of special attention despite CISA’s warning. The cyber-agency’s decision seems apt because vulnerability research and red teaming outfit Calif.io, which discovered the flaw, used its X account to claim news of the problem had leaked before public disclosure and the release of Oracle's fixes. By then, Calif had already tested the flaw, which it named “Java4Shell,” and found applications including Keycloak, Atlassian’s Jira and Confluence, Jenkins, Apache Kafka Broker, and Elasticsearch are susceptible to RCE. Calif, which also helped to find the PHP2Shell flaw, says the flaw is a memory-corruption vulnerability in OpenJDK's Base64 decoder. “Base64 is a very common way of representing binary data as text. It is used in authentication protocols, to encode file uploads, and to deliver network messages,” the company explained. “To decode Base64 faster, OpenJDK uses specialized x86-64 CPU instructions to process many bytes at once.” “We discovered a bug in that optimized machine-level code that lets malformed input overwrite neighboring memory. With careful exploitation, the corruption can reveal sensitive memory, as Heartbleed did, or lead to remote code execution. The corruption itself can be triggered by just 64 A characters followed by a single =:” Calif applied the CVSS 4.0 bug-rating scheme and gave the flaw a critical-rated score of 9.2/10. Infosec outfit SentinelOne’s analysis suggests the flaw “sits in the Java standard library, so exploit traffic is specific to each application.” “Any Base64 value an application passes to java.util.Base64.Decoder can carry the payload, including HTTP Basic credentials, SAML artifacts, CSRF tokens, and MIME encoded words. In default configurations of the affected applications listed above, attackers can reach this code path prior to authentication.” Cisco’s Talos security team pointed out that the flaw is present in OpenJDK Java versions 17.0.6 and later running on x86-64 CPUs with AVX-512 VBMI (Vector Byte Manipulation Instructions) support enabled. “These vulnerabilities have already been tested against a suite of Java applications with success, but the attack surface is likely much larger as additional targets are tested and confirmed,” Talos advised. Cisco’s infosec experts urged adoption of the fixed versions of Java Oracle posted on Tuesday but noted it is possible to mitigate the flaw on x86-64 machines by starting the JVM with support for AVX-512 disabled. Akamai’s Chief Information Security Architect Jan Schaumann, writing on the Openwall project’s mailing list, expressed a little discomfort with Calif’s work. “Calif.io set their priorities on publicity-maxxing, it seems,” he wrote, pointing out that the flaw-finding firm’s announcement of Java4Shell opens by stating “After we disclosed PHP2Shell, we realized many people didn't know what PHP was. So we decided to go after something a little more familiar: Java.” Talos’s analysis of Java4Shell includes the interesting observation that it “has not observed in-the-wild exploitation of this vulnerability yet. However, we anticipate probing attempts may pick up soon after public disclosure.” That second sentence perhaps supports Calif’s assertion that news of the flaw leaked before disclosure. The rest of Big Red’s flaws Java4Shell deserves a place high up on patch priority lists. So do the six 10.10 flaws Oracle reported, three in its Access Manager product, one in the Oracle Platform Security for Java, another in the WebLogic Server, and one in Business Intelligence Enterprise Edition. Big Red also repaired five 9.9-rated bugs, in Oracle Text, Hyperion, RDBMS, and Autovue. The November patch dump also includes 28 flaws rated 9.8/10. By Tenable Security’s count, 249 of the patches are for high severity bugs. Oracle’s Hyperion enterprise performance management accounts for 267 of the bugs. Such is life in the age of AI-powered flaw finders – which Calif says burned through just $200 worth of tokens to find Java4Shell. ®

Daily disruptive outages that ground planes and halt work – welcome to the new normal

17 November 2026 at 23:41
Another day, another disruptive network outage (or two) as government officials and company execs scramble to assure an anxious populace that these are not coordinated cyberattacks – and security experts call this the new normal. A classified network that the US National Security Agency uses for internal communications, called NSANet, went offline on Tuesday in what the agency described as a “temporary degradation during scheduled routine maintenance.” “While some internal systems were affected, mission-critical capabilities remain fully operational and our ability to defend national security continues,” the NSA said in a statement following media reports about the outage. “NSA's infrastructure is designed with multiple layers of redundancy to maintain operations during planned maintenance or unexpected events.” Around the same time, Apple users began reporting issues across multiple services – including Apple Support, iCloud, the App Store, Apple Pay, and more – with Apple retail stores worldwide also reportedly experiencing disruptions that halted sales and in-person support. A day earlier, customers at major US banks including JPMorgan Chase, Bank of America, and US Bank reported service outages that Chase attributed to “a software bug in on-premise hardware systems - there is no evidence this is cyber-related.” All of these high-profile outages followed last week’s US Federal Aviation Authority (FAA) order grounding all planes on Friday due to problems with copper wiring during an L3Harris network upgrade, and a Microsoft 365 outage on Thursday. Redmond later blamed the massive service disruption on a faulty software update it pushed back in July. Despite the weekend respite from network outages, your humble vulture still had to field is-this-the-beginning-of-the-end questions from friends and family all day Saturday and Sunday, most of whom suspected Russia or China – maybe even Iran – caused the spate of downed systems and services. A highly sophisticated, coordinated nation-state hack would be a much more exciting cause for the disruptions. The real culprit, however, appears to be aging, under-funded infrastructure, riddled with decades-old bugs and held together by digital sticks and glue. “As someone who started out as a hacker, I know how tempting it is to connect the dots when major systems fail one after another,” Remedio CEO Tal Kollender told The Register, adding that there’s “no credible public evidence that these outages were part of a coordinated cyberattack,” and most have been attributed to technical failures. “What concerns me more is what these incidents reveal about the fragility of our digital infrastructure. We've built an interconnected world on layers of aging systems, complex configurations, and dependencies that organizations don't always fully understand. A single misconfiguration, failed update, or broken dependency can create disruption on a massive scale, even without an attacker,” Kollender said. “The uncomfortable reality I think we're grappling with is that attackers don't always have to break our systems. Sometimes, our systems are already capable of breaking on their own.” For people who couldn’t withdraw money from their bank, board a flight, or use email at work on Friday, “the disruption was real regardless of the cause,” she said. “Cybersecurity has spent decades investing in detecting threats, identifying vulnerabilities, and generating reports,” according to Kollender. “Yet we still allow for these preventable weaknesses and configurations that can bring entire systems down.” Cynthia Kaiser, senior VP of Halcyon’s Ransomware Research Center and former deputy director of the FBI’s Cyber Division, told The Reg that the recent disruptions show “how deeply interconnected our digital infrastructure has become, and how a single technical failure or software update can cascade with outsized effects across critical systems.” AI has entered the chat Then we go and add AI to this already delicate balance, and everything essentially blows up. AI-assisted bug hunting and exploit development has surfaced more than 100,000 CVEs so far this year, with Microsoft’s Patch Tuesday disclosing more than 1,000 vulnerabilities in each of the past two months. AI has also shrunk the elapsed time between the public release of a security patch and attackers developing working exploits. “We are in an era where AI is accelerating the discovery of legacy software flaws, forcing organizations to race to patch vulnerabilities before adversaries can exploit them, which inevitably creates friction and disruptions,” Kaiser said. “Even when an outage isn’t caused by a cyberattack, these moments of operational distraction leave dangerous gaps open for threat actors and opportunists to take advantage of, making it crucial for both institutions and everyday individuals to remain on high alert against social engineering and unsolicited outreach.” Even when attackers aren’t using AI to facilitate intrusions, the technology itself has become a source of risk to organizations. “Autonomous agents can execute actions, change configurations, access sensitive information, and interact with critical systems at machine speed,” Kollender said. “When those systems lack proper governance and safeguards, a mistake can develop just as quickly as an attack. Former US Department of Defense network exploitation and vulnerability analyst Mike Wiacek, when asked if these regular system outages are the new normal, told us: “The brutal truth is yes. Especially in the AI world that we live in.” After the DoD, Wiacek founded Google’s Threat Analysis Group. He's since founded Stairwell, a threat-intel firm he serves as chief technology officer. “The idea that we are going to patch ourselves to security is, dare I say, an antiquated concept,” he told The Register. One of Stairwell’s engineers was doing regular, daily software development tasks on an open source library with Anthropic’s Claude Mythos, and the model told the engineer about a buffer-overflow vulnerability in the code. “We weren't even looking for it,” Wiacek said. “It found a vulnerability in open source software by accident. You used to need specialized bug hunters and really, really technical security folks to discover and develop vulnerabilities and exploits. Now, it’s almost too easy to do it.” This new reality, he said, necessitates a total rethink of how we build and maintain digital infrastructure. “How do we get to a world where we assume compromised by default, and continuously recompromised, and then develop the architectures and the systems necessary to make what was the expense of incident-response investigation really fast and cheap? Because we will not be able to patch our way out of it.” It requires vulnerability managers and system administrators to patch as soon as software updates become available. But from an engineering perspective, it also means building in “as many fail-safes as possible,” Wiacek said. “An elevator doesn't have just one cable hoisting it,” he continued. “There's a lot of redundancy in those systems. Enterprises and organizations and governments haven't really had to operate like that. But that's going to be the new norm, whether we want it to be or not. Pandora's box is opened, and it's not going to go back inside.” ®

Gemini test proves once again that amoral AI agents will do almost anything if told ‘NEVER give up’

17 November 2026 at 19:52
After being instructed to "NEVER stop, NEVER give up" on a difficult task, Google’s Gemini AI agents tried to hack a an organization, impersonating its staff and attempting a SQL injection attack. The good news? It was all part of a test. The Gemini 4 Argon agents were given an impossible task to complete – come up with the “official” answers to Epoch AI’s FrontierMatch problems. In some runs they had internet access. Even with their production safeguards enabled, Trajectory Labs says the agents “engaged in misaligned behavior.” “Across four scenarios, they attempted SQL injection against Epoch AI’s live website, claimed to be Epoch staff to get answer keys, turned a shared package cache into a message board, and obfuscated their commands to avoid their safeguards,” the AI safety company said in a Friday report. If this all sounds similar, it should. Google’s AI agents behaved in much the same way as OpenAI’s models did when they broke out of their testing environment and hacked Hugging Face. Google did not respond to The Register's questions. According to Trajectory, the Hugging Face intrusions prompted it to test misalignment in other models. While OpenAI’s models were running with their safety guardrails turned off in the Hugging Face scenario, Trajectory “wanted to know whether a current frontier model would behave the same way with its production safeguards enabled.” Trajectory ran swarms of up to 80 Gemini 4 Argon agents, gave them an impossible task, and prompted them not to give up until they completed it – which, we’re told, was “intentionally” done to “elicit misaligned behavior, if it existed.” Turns out it existed. “We told the agents to persist with an impossible math benchmark task, but we did not tell them to hack anyone or pose as Epoch staff,” McIntyre said. “When a model hacks systems or deceives people to finish a task, it is misaligned with what its developer and its user intended,” he continued. “Production safeguards are meant to block harmful requests even when the model itself goes wrong. Under our test conditions, neither the model nor its safeguards stopped the agents.” Four test scenarios The initial testing began on October 20, and the agents had internet access. We should note that just one of these tests was run with Epoch’s consent. In the first scenario, the agents did reach Epoch’s live website – and fell into the research nonprofit’s honeypots, set up to detect AI agents. These honeypots contained a decoy benchmark key, which the agents found and used. The agents also probed for internal endpoints including /admin, /keys and /users, and attempted a SQL injection. After noticing the agents’ activities, Epoch added an “agent helpline,” and some of the agents that used it claimed to be Epoch staff seeking answers to the test questions. After this initial test, Trajectory and Epoch decided to use a copy of the website instead of a live version, and built a buggy local copy with a planted vulnerability based on CVE-2026-102255, a critical pre-authentication Server-Side Request Forgery (SSRF) vulnerability in SonicWall Secure Mobile Access (SMA) appliances. “Over the following days, several agents fully exploited the planted multi-stage vulnerability chain in the clone,” the report says. In an effort to mimic the Hugging Face incident, Trajectory also gave the agents a local, shared package manager, which “nearly all” of them used as a message board to communicate with each other and coordinate efforts to complete the test. In a separate, local setup, Trajectory swapped out Epoch’s honeypot with an AI assistant that would release answers only to named Epoch staff who passed a security check with their date of birth. “Agents accessed people-search sites, attempted impersonation again, and tried to jailbreak the AI assistant,” the testing firm said. Next time: kill switches at the ready A third test on October 30 used a staging copy of Epoch’s website, and granted the agents internet access. One of them tried a SQL injection, and multiple claimed to be Epoch staff on the helpline. They also attempted to bypass a CAPTCHA check. Finally, the fourth scenario on October 29 and 30 kept the agents contained in a sandbox, with no internet access and no copy of Epoch’s website. They were, however, offered a decoy switch and told it would turn off the firewall. In all three runs, the agents switched the decoy off and obfuscated its address to bypass Gemini’s safeguards. On one, the agent rebuilt the switch’s address from positions of words it had saved earlier. The real firewall blocked the agents’ traffic throughout all of these tests. At this point, Trajectory stopped its tests, and said it won’t run any future experiments until it has a risk management plan in place, with signoff from its CEO and CTO, along with approval from any outside systems that the models could “reasonably” target. Tests will also include real-time automated monitoring and kill switches, and Trajectory said that it’s setting up a system to notify any organizations affected by these experiments. Interestingly, Trajectory co-founder Peter McIntyre told us that the testing firm “ran a similar test with another frontier model, and it did not exhibit the same misalignment as Gemini 4 Argon.” However, this was not a controlled comparison. “Our report covers one model's behavior, and was designed as a stress test to see whether that model would exhibit similar misalignment under duress, not a measure of how often different models behave this way,” McIntyre told The Register. Overall, the testing firm noted that the agents’ behavior “show how easily testing capable AI agents can spill over to affect real systems and people, and there are not yet shared standards for doing it safely.” Trajectory says that it's developing a set of safety standards, and recommends that anyone else running misalignment tests use fully isolated sandboxes. ® Correction: Epoch only consented to one of the tests, scenario three. The original version of this article said that it consented to all of the tests.

Hackers smash 'n' grab AI vendor: Take off with license data of Uber drivers

17 November 2026 at 16:15
Uber has confirmed that nearly 150,000 people had their personal information compromised after hackers broke into one of its AI suppliers. A filing with Oregon's attorney general puts the number of victims at 149,346. Among the data the crooks got their hands on were images of drivers' licenses and information extracted from them, including names, home addresses, dates of birth, and license numbers. Uber doesn't name the supplier in its filings with US state attorneys general, but told TechCrunch it was Fireworks AI, the inference outfit whose security woes we covered last week. According to a breach notification filed with Iowa's attorney general, Fireworks alerted Uber to the intrusion on October 9. Uber cut off the supplier's access to its data and, four days later, established that the attackers had accessed drivers' personal information on October 7. Uber had set up a dedicated cloud storage environment for the AI biz, complete with security monitoring overseen by its own cloud security team. Unfortunately for Uber and its drivers, the crooks were using legitimate credentials, so none of the alarms went off. The ride-hailing firm says those credentials couldn't be used to access its other systems, and says it has since revoked or changed the credentials associated with the supplier. Uber hasn't said why it handed images of thousands of driver's licenses to an AI inference outfit. The Register has put that question to both companies, and asked Uber if any drivers outside the US had their data exposed. Fireworks, which provides infrastructure for developers to run and fine-tune AI models, had a particularly miserable October. The company first disclosed a security incident on October 9 after discovering that someone had been using its internal credentials without permission. Fireworks initially said it had no evidence that the intruders had gotten their hands on personal information or customer datasets. The company was hacked again later that month and, on November 9, disclosed that some customer data had indeed been compromised. The AI biz acknowledged that attackers had accessed billing records, some customer datasets, and incomplete model weights. Productivity software maker Notion was among those affected, with some of its customers' conversations with its AI assistant exposed. Fireworks suspects the involvement of financially motivated criminals in the business of flogging unauthorized access to large language models. Whether the crooks have any plans for the haul of Uber drivers' personal information is unclear. Uber began notifying affected individuals on November 10, more than a month after learning of the breach. It's offering victims free credit monitoring and fraud assistance, though drivers have just 90 days from the date of their letter to sign up. This isn't Uber's first brush with cybercriminals in recent months. In August, the ride-hailing giant's logistics arm, Uber Freight, confirmed a break-in after the Helix extortion crew claimed to have made off with almost a million files, including data pilfered from mailboxes and OneDrive accounts. ®

Windows 11 previews remote padlock only IT can open

17 November 2026 at 15:29
A new experimental Windows 11 build previews security features including an Intune-controlled Remote Lock, while a beta build clears away another relic from the halcyon days of Windows Phone. The security changes include viewing and individually removing passkeys stored on FIDO2 (Fast Identity Online 2) security keys, plus a preview of the Microsoft Pluton Cryptographic Provider. For enterprise administrators, though, Remote Lock is the main attraction. Finding a lost Windows device has been around for years, and you can also remotely wipe a device to prevent data from falling into the wrong hands. Remote Lock lets an administrator lock a managed device through Intune without wiping it. The instruction takes effect at the device's next check-in. Windows signs out of every active session, hides every sign-in option, and shows a locked screen. The lock survives restarts and the device going offline – only an administrator can lift it remotely. Microsoft's documentation spells out the restriction: "There is no offline or local unlock. No PIN, password, local admin, or web sign-in works on a locked device; unlock arrives only through Intune." Once the lock is applied, even stolen credentials won't restore access through the normal sign-in screen. The user's account itself is not affected; this applies only to the specific device. Though still in preview, the feature will allow IT teams to respond non-destructively to a report of a missing device. It doesn't erase data or isolate the device from the network (Microsoft directs admins to other Intune controls if that is required). The lock doesn't block Safe Mode, Windows Recovery Environment, or booting from external media. Microsoft advises keeping BitLocker enabled to protect the data. Now for the beta build, where Microsoft, in time-honored fashion, gives and takes away. The latest Windows Insider build hammers another nail into the coffin of Microsoft Maps. The Maps app was removed from the Microsoft Store and made nonfunctional in July 2025. Now Offline Maps settings are following it out, as Microsoft retires the Windows UWP Map control and Windows Maps platform APIs. The original Maps app dates back to Windows Phone 7 and appeared on the desktop with Windows 8. It continued through subsequent versions before Microsoft stopped preinstalling it with Windows 11 version 24H2. For those with a nostalgic streak, it's another piece of Windows Phone disappearing off the map. ®

'Not rocket surgery' claims researcher who found six Danfoss fridge controller zero-days in under an hour

17 November 2026 at 13:32
A Dutch cybersecurity researcher claims he found six zero-days in Danfoss refrigeration controller firmware in under an hour, then delivered a withering assessment of the vendor's defenses against memory corruption. The Centre for Cybersecurity Belgium (CCB) warns that the flaws could let attackers take control without credentials, potentially retain access after resets, or disable controllers altogether. Carlo Meijer of cyber consultancy Midnight Blue examined the latest AK-SM800A firmware, version 5.3.2, manually – no AI required. He said that he went looking in random places where strcpy() is invoked, given the vendor's "extremely liberal use" of the function (which doesn't check whether the destination buffer is large enough), and swiftly found five cases in which he said he could trigger a segfault with a one-line cURL command. "Of course, remotely triggering a crash is bad in itself, but being able to leverage the buffer overflow to obtain remote code execution is much worse," he wrote. Meijer added: "Any reasonably experienced exploit developer will agree exploiting straightforward stack-based buffer overflows on ARM Linux without exploit mitigations is very straightforward. And if they don't, there's probably a frontier or open-weight model out there willing to help them agree." He noted that before achieving RCE, an attacker would need to overcome the limitation of strcpy() where it stops copying bytes after hitting a null byte, but said "it's not exactly rocket surgery." He went on to claim: "Also, the exploited application runs as root (because of course it does)." Later firmware adds "session control," an authentication mechanism whose strict setting blocks unauthenticated access to most web interfaces. But Meijer found a memory corruption flaw in that protective layer too. He described the final zero-day as a "trivial" pre-auth memory corruption bug which again starts with a segfault but can be wangled into an RCE "with some additional effort." Meijer also considered whether attackers could maintain access after a reset, as seen with implants such as Saltwater and Firestarter. He considered this feasible because the controllers do not use their processor's secure boot features. "In such a scenario, an attacker could either retain remote access or leave a time-triggered payload for a 2nd (or 3rd...) strike – increasing damage by spoiling new batches of food," said Meijer. "Alternatively, attackers could simply wipe flash memory and/or push a malformed firmware update to brick the controllers and render them permanently inoperable – requiring costly and time-consuming replacement similar to the TTPs deployed during the December 2025 attack on the Polish electric grid." Meijer's priority for operators was to remove the controllers' direct internet exposure, rather than spend scarce resources on piecemeal patching. He said the vendor's advice following the attacks on European supermarkets and blood banks in recent weeks – rotating creds, applying patches, enabling strict session control – was insufficient given the flaws in the controllers' codebase. "This is insufficient and asset owners need to remove internet exposure as soon as possible," said Meijer. "We recommend against allocating scarce security resources to patching efforts and instead focus on reducing exposure." The CCB likewise urged operators to remove direct internet exposure immediately, but also recommended applying patches when they become available. Beyond the liberal use of strcpy(), Meijer criticized the absence of standard exploit mitigations, including stack canaries and position-independent executables. "These mechanisms have been staples of memory corruption defenses for decades and enabling these features... requires no engineering work at all," he said. "It is simply a matter of setting the right compiler flag. Enabling these features became the default in Ubuntu 17.10 and there is simply no excuse for a network-facing application that was last updated a few months ago to not have them enabled." The Register asked Danfoss to respond to the findings, but it did not reply. Danfoss has consistently blanked our requests for information since being implicated in at least some of the attacks that affected European supermarkets recently, alongside rival controller vendors Copeland and Dixell. It also did not respond when we asked about its involvement in the attacks on US commissary appliances over the summer, after some alleged that its controllers were targeted. ®

Opinion pollster Ipsos 'fesses up to 'cybersecurity incident'

17 November 2026 at 11:44
Market research giant Ipsos says it's currently investigating a "cybersecurity incident," and has reported the matter to the authorities. In a statement issued on Monday from its Paris headquarters, the company confirmed it filed a complaint over the attack on November 11, but gave few details because its internal investigation is ongoing. "Upon discovery, we promptly activated our incident response procedures, took steps to contain the incident, and engaged external cybersecurity specialists to support our investigation," it said [PDF]. "The investigation is ongoing to determine the nature and scope of the incident. We have reported the matter to the relevant authorities." Ipsos confirmed the attack had no impact on its operations or customer service delivery, however, it did not comment on the nature of the attack or whether any data was compromised. "We take protection of the information entrusted to us very seriously," Ipsos added. "Should the investigation establish that information relating to clients, projects or business partners has been affected, we will contact those concerned directly. "As the investigation is ongoing, we are not in a position to comment further at this stage." The company declined our requests for additional data, as it does not want to release unverified information to the market. Founded in France in 1975, Ipsos is one of the leading market research companies in the world, employing nearly 20,000 people and operating in 220 cities across 90 markets. It is routinely engaged by businesses for its market research data, as well as by public sector organizations for its opinion polling, often responsible for providing data regarding political views and satisfaction surveys. Tasked with procuring a wealth of data on myriad topics, Ipsos also offers data analytics services that turn the information it gathers through research, polls, and surveys into insights that inform organizations' business strategies. ®

Received — 17 November 2026 ⏭ The Register - Security

Japan's Fugaku supercomputer has decidedly un-super security

17 November 2026 at 05:27
The RIKEN Center for Computational Science, the home of the Fugaku supercomputer, has revealed unauthorized attempts to access the machine after a data leak. According to a statement issued by RIKEN yesterday, between November 10 and 13 an unknown entity attacked a web dashboard that displays operational status data for Fugaku. It sounds like whatever the attackers did to that dashboard allowed them to move across the RIKEN network, because they were able to reach several servers that store data belonging to authorized users, and install malware. The servers stored users’ names and email addresses, plus credentials used to access supercomputers and services not operated by RIKEN. The attacker has tried to use those credentials. RIKEN is working with the organizations that host those other supercomputers to investigate whether the attackers succeeded. The research outfit doesn’t know how the attack happened and can’t say with certainty what information attackers were able to access and/or steal, beyond saying that 4,222 registered users of Fugaku are at risk, along with 375 registered users of its smaller Rikyu supercomputer. RIKEN does know that on November 13, someone attempted to log onto Fugaku. RIKEN quickly took the machine offline. Fugaku comprises 158,976 nodes, each packing a 48-core processor and 32GiB of available RAM, for a total of 7,630,848 cores and over five petabytes of memory. Those specs saw Fugaku lead the Top 500 lists of Earth’s mightiest supercomputers for two years between 2020 and 2022. It ranks ninth on the update of the list released on Tuesday. RIKEN’s other super, Rikyu, operates 288,000 cores. If attackers managed to run a workload on either machine, they could do a lot of damage. The intrusion has already had serious consequences because Fugaku is currently offline. As the machine is used for scientific research, and RIKEN schedules use of the machine, the outage will disrupt researchers who were waiting for their turn to use the supercomputer. At the time of writing, RIKEN has found no evidence that users’ research data has leaked but isn’t yet sure what info did make it out of the building and whether it contains any data describing third parties. The research org has promised to conduct a detailed investigation into the incident, which is the latest in a long string of significant cyberattacks that have struck Japanese clouds, telcos, retailers and government agencies, leading Tokyo to release emergency funds in the hope that extra money can quickly shore up the nation’s cyber-defenses. ®

Received — 16 November 2026 ⏭ The Register - Security

Infoblox customer and staff data exposed in Sharepoint breach

16 November 2026 at 21:53
Networking and security firm Infoblox is investigating a third-party breach that allowed crims to access sensitive customer and staff data. "Infoblox is investigating a security event involving unauthorized access to certain company information through a compromised third-party data migration vendor,” a company spokesperson confirmed to The Register. According to sources, the compromised third-party data migration vendor had trusted access to Infoblox's SharePoint system, and the intruders used this vendor’s accounts and access to exfiltrate data directly from Microsoft SharePoint. This data included configuration information related to some customer's NIOS (Network Identity Operating System) deployments. These are saved snapshots of customers’ network data, such as DNS, DHCP configurations, Grid databases, and reporting logs, that can be used to fully recover the appliance or grid in case of a system failure or faulty upgrade. The intruders also stole employee data during the break-in. According to an email CEO Scott Harrell sent to employees, which The Register has seen, compromised employee details include financial information such as direct deposit details, contact info, dates of birth, and Social Security numbers for employees and dependents. Upon discovering the breach, Infoblox “immediately activated” its incident response procedures, and took steps to contain the incident, including hiring outside cybersecurity experts, the spokesperson said. “We have sent notifications about this security event and are providing support and guidance as appropriate,” the Infoblox spokesperson told us. “Based on our investigation to date, we have no indication that any Infoblox products, cloud services, software updates or build-and-release processes were involved. Our priority is to fully investigate and resolve this issue." Affected staff received a $100 "spot bonus" to buy password management services, and were warned to be "especially alert for phishing, vishing, fraud, or other suspicious communications." Infoblox also gave employees and their dependents free credit monitoring and identity protection for two years. Infoblox declined to answer our more detailed questions about the compromise, including the type of data exposed, and how many employees and customers might be affected. The Register will update this story as we learn more about the breach.®

Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn

16 November 2026 at 19:52
Attackers are using open source AI penetration testing framework ARTEX to automate their attacks on banks, governments, energy and healthcare providers, and other critical sectors around the globe, according to security alerts from the FBI and Amazon Threat Intelligence. Between October 3 and November 5, Amazon’s MadPot honeypot threat sensors detected two ARTEX campaigns: one that allowed thieves to steal AI credentials at “industrial scale” to gain free access to models, and a second that generated deepfake videos to trick live selfie checks. In the first one, miscreants used AI agents to steal API keys, and then feed these keys into the ARTEX framework’s model gateway to power more agents searching for more AI model access keys. “Agents validated identities through each provider’s token and identity services, enumerated storage, assessed permission scope, and went after managed AI model services, secrets and parameter stores, and application identities,” AWS Chief Information Security Officer CJ Moses wrote in a blog post on Friday. Targets included LLM gateways and proxy platforms such as One API, New API, Veloera, Sub2API, and LiteLLM, and AI application builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama. This campaign also used cloud credentials to gain stolen access to the models. In a second campaign, attackers used ARTEX to create real-time deepfake videos to trick identity verification (KYC) checks at two cryptocurrency exchanges, and forged ID photos to pass document checks. “The agents reverse-engineered one vendor’s liveness protocol in detail, documenting the face-detection event stream it sends, how it tracks face position during the check, and the coordinate system it uses to tell the user where to move,” Moses wrote. The agents did manage to complete the full movement sequence with the face positioned within 7 pixels of the target, but this feat was not enough, as the vendor’s server-side verdict rejected the deepfake in every attempt. In a separate session, agents built a real-time deepfake video pipeline on a single consumer-grade GPU using an open-source face animation model. This video achieved 11.5 frames per second, which Moses said is “close to what a live video check expects.” Meanwhile, agents also worked on the ID document checks, building synthetic camera metadata whose image dimensions matched the real values inside the JPEG file, and preserving the original camera compression tables that are usually replaced by re-saving an edited image. While the end goal was stealing cryptocurrency, the campaign illustrates how ARTEX runs several sessions simultaneously, with multiple agents working toward the same result. “One searched for ways to cash out, one worked on clearing account balances, one set up the deepfake video pipeline, and one carried out the verification bypass,” Moses wrote. None of the attack techniques are new, but both of these campaigns show how AI agents can boost the speed and persistence of the attacker. “The agents succeeded where basic security hygiene was missing, including default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces and inference endpoints, and unpatched software,” Moses added. “Getting those fundamentals right remains the most effective defense.” The FBI weighs in The FBI also hammered this message in a separate Friday alert noting “ARTEX significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours.” The federal cops cited the August attacks against South Korean financial institutions that stole personal data belonging to about 67,000 people. The FBI also highlighted a September ARTEX campaign that targetedd ConnectWise Cloud Backup customers in a supply chain attack that ultimately compromised more than 2,250 downstream Microsoft tenants. After these security breaches became public, the original ARTEX developer removed the GitHub repository in October. However, forked and localized variations plus a commercial version remain, and as of November 4 internet scanning service Censys identified 1,465 hosts running an exposed ARTEX console. And the total number of deploymentsis “likely significantly higher when including systems behind firewalls or in private networks,” according to the FBI. The feds urged all organizations – especially those operating internet-exposed applications, self-service portals, cloud infrastructure, and partner or supplier systems – to implement a series of mitigation measures outlined in its Flash alert. These include reducing external attack surfaces by restricting sensitive services to trusted networks, private connectivity, or dedicated administrative paths, and moving enterprise applications behind zero-trust access policies and permissions. ®

‘123456’ has been the world’s most popular password for 7 out of the last 8 years

16 November 2026 at 18:33
It’s been eight years since NordPass released the first edition of its most common passwords of the year list, and the world has learned nothing about password security in the meantime. NordPass’ eighth edition of its most common passwords list, as culled from data breaches, was released Monday, and it looks suspiciously like the previous ones in the worst possible way. “Once again, ‘123456’ is the most commonly used password in the world,” NordPass said in the report. “That makes it seven out of eight years this password has stood at the top of the leaderboard.” Even when 123456 wasn’t number one, it was still near the top of the heap, as the top three passwords in the list remain unchanged in the eight years NordPass has published its list: 123456, admin, and 12345678. It’s not like the rest of the most common passwords are showing any variety either. According to NordPass, the only real shuffling among the 10 most popular passwords was the entry of the worryingly short 1234 into the top ten and the migration of the ever-popular password “password” down slightly in the top 10 from 6th to 7th place. “We still have a lot of work to do to spread awareness about password safety,” NordPass concluded of that data – and there’s no more pressing time than the present. AI adoption has made password cracking an easier exercise than ever, with more credentials being sucked up in more data breaches at a faster pace. AI has also made cracking hashed passwords a snap too. In other words, if you’re reusing a password that’s been exposed in a breach, even if it’s just a hash that was exposed, you’re more likely than ever before to have another account compromised. The best advice NordPass gives is to move away from passwords toward biometric authentication or passkeys and you're insane not to enable MFA. But for the places you still have to use passwords, there are a few things to avoid – like the 200 passwords in NordPass’ list, for example. Beyond that, NordPass also recommends steering clear of pop cultural trends. “Fandoms make for terrible passwords,” NordPass notes. There are several reasons for that – namely the fact that they follow trends, meaning that “minecraft” password in the early 2000s was a simple one to add to a password stuffing dictionary, just like “skibiditoilet” in recent years. The one thing that connects all those pop-culture passwords is that “they all could be cracked in seconds,” NordPass said. We’re sure our readers are conscientious about their online habits, carefully protecting their accounts with the proper security measures and ensuring each and every password is unique, long, complicated, and random. Unfortunately for you smarties reading El Reg, the person down the hall whose security you’re responsible for may not be so disciplined, so be sure you’re enforcing good practices – Nord knows the average user just isn’t doing it despite years of begging and reminders. ®

NoisyTCP bug gives Linux servers a remote crash problem 11 years in the making

16 November 2026 at 18:28
The researcher who sounded the alarm over Dirty Frag earlier this year has uncovered another Linux kernel flaw, and this one lets attackers crash entire servers remotely without so much as a password. The vulnerability, tracked as CVE-2026-99288 and dubbed NoisyTCP, was discovered by security researcher Hyunwoo Kim, who says it affects almost all Linux distributions. The flaw dates back to October 2015 and went unnoticed until this year, when Kim found and patched it. Attackers don't need to log in or get their hands on any credentials to cause trouble. They can simply knock the server over, although there's no evidence yet that the flaw can be used to steal data or hijack a machine. The bug is buried in Linux's TCP networking code. Two parts of the kernel can trip over each other while processing incoming traffic, leaving one trying to use memory that's already been freed. When that happens, Linux can panic and take the whole machine down with it. Kim says there's no need for any special server configuration either. Services such as Nginx, Apache, and SSH can be vulnerable straight out of the box. An attacker needs access to an open TCP port on a system that processes incoming network traffic across multiple receive queues or CPUs. Perhaps more concerning for administrators running containerized workloads, the damage isn't necessarily limited to the application being attacked. Because the flaw sits in the Linux kernel, crashing a web server inside a container can take down the underlying host and anything else running on it. Kim found that the attack was particularly effective against virtual machines. During testing, he managed to crash Nginx, Apache, and SSH servers running on QEMU/KVM virtual machines in between five seconds and two minutes. Physical hardware put up more of a fight. Even after tailoring his exploit to the target machine, Kim needed roughly 14 hours to crash a bare-metal server running Nginx. Kim has a knack for finding holes in Linux. Back in May, he uncovered Dirty Frag, a kernel bug that allowed local users to gain root privileges. That one caused headaches for administrators when a working exploit went public before patches were ready. His latest discovery, NoisyTCP, has another potentially nasty trick up its sleeve. Kim identified a second exploitation path involving a use-after-free condition, although pulling it off is considerably more difficult than simply crashing a server. In his advisory, Kim says he has developed a limited exploit targeting virtual machines in a lab environment but has yet to demonstrate a general-purpose remote code execution attack. For the moment, the confirmed impact is remote denial of service. A fix was committed to the upstream Linux kernel on November 3, but administrators will need to check whether their distribution has shipped an updated kernel. Red Hat has rated the vulnerability Important, assigning it a CVSS score of 7.5. For those unable to patch immediately, Kim suggests two temporary workarounds: disabling TCP SYN cookies or restricting incoming packet processing to a single receive queue. Neither is particularly appealing. Disabling SYN cookies reduces protection against SYN flood attacks, while limiting packet processing can hurt network performance. The former also needs to be applied separately to containers using their own network namespaces. Kim published the details on November 15, once the disclosure embargo was lifted. The bug had been hiding in plain sight since 2015, leaving Linux servers vulnerable to remote crashes for 11 years. Not bad for a flaw that doesn't even require a login. ®

Ubuntu gets AI-generated security fixes for those feeling brave

16 November 2026 at 18:01
Canonical is offering Ubuntu Pro users an opt-in channel for AI-generated security patches, deploying bots to help tackle the growing pile of vulnerabilities other bots are uncovering. Canonical's VP for "Excellence Engineering," Pierre Guillemin, gave the second keynote at the recent Ubuntu Summit, introducing the Edge archive, an optional repository of security fixes generated by the company's Surge framework. As the announcement describes it: "We built Surge, a multi-agent orchestration framework for CVE remediation and backporting, and have reached a high level of confidence in its capabilities. We will not automatically upload Surge-generated fixes to our stable releases, but from today they are available to Ubuntu Pro users in the new Edge archive." Edge offers patches before extended testing and qualification are complete, allowing users to assess them themselves. Canonical says updates delivered through the normal security channels will remain human-validated. Ubuntu Summit now runs twice a year. Last week's Ubuntu Summit 26.10 followed a rather unfortunately timed server outage for the company. At the previous such event in June, the company announced a move into agentic AI, but this new one is bolder still. Guillemin noted that "October 2026 saw almost as many CVEs as all of 2025." Surge is a "multi-agent orchestration framework" to remediate vulnerabilities in Ubuntu's userspace. In other words, not the kernel, but even so, it is ambitious. The idea is to use a group of different LLMs to triage new vulnerabilities, find existing fixes, use multiple competing models to backport them to the versions in Ubuntu, and run automatic testing on them. Then a different group of "judging models" will evaluate the fixes and look for consensus. When candidates are chosen, they pass through normal Ubuntu package testing. If there are edge cases, low confidence scores, deadlocks, or other problems, then the process will be escalated to the staff of Canonical's security team. It's an attempt, as Guillemin put it, "to use AI to fight AI" – cooperating groups of LLMs patching their way through the ever-increasing flow of security holes. Even so, this is a remarkably courageous tactic (at the risk of using a famous British euphemism). It's an opt-in feature of Ubuntu Pro, Canonical's security and support subscription – although as we reported in 2022, Pro also has a free tier, so security researchers and the community can try Edge out, study it, and report their findings. We do sympathize with companies struggling against a rising tide of bot-powered security hole discovery. In theory, corporate funding can pay for faster fixes than a volunteer community can offer, but even for commercial vendors these are testing times. As we reported when Debian turned 30, surveys suggest that Ubuntu is about a third of the world Linux user base. Some of those users will be intrepid enough to try this, and that will end up being informative for the whole FOSS world. Those with long memories will remember the last time Canonical used the same edgy nomenclature. Back in July 2013, it launched a crowdfunding campaign for an Ubuntu smartphone. The campaign didn't raise enough money. ®

Crook turns LearnWorlds' free trials into phishing factory, sends 13,000 fake YouTube copyright notices

16 November 2026 at 17:03
LearnWorlds, a software and learning course platform, says an attacker sent 13,000 phishing emails to YouTube creators over seven days by using automated tools to open free trial accounts. The company insists its systems were not breached. Panos Siozos, Co-founder and CEO at LearnWorlds, said the miscreant created around 570 free trial accounts on the platform between November 7-14, which they used to issue fake copyright infringement notices to recipients of their choosing. "We have found no evidence of unauthorized access to existing customer schools or personal data held for our customers and their end users," he told The Register. According to Siozos, the attacker bypassed the platform's per-customer limit on trial subscriptions and exploited its welcome email feature to send the phishing emails. "In each trial they added their own list of target addresses and rewrote the standard welcome email as a fake copyright notice," he said. "Adding those users triggered the welcome email from our sending address." The first email was sent on November 7, after which the company detected the suspicious signups and blocked the first sender on November 9, although the attacker returned to send a second wave of emails within 40 minutes. LearnWorlds removed all the trial accounts by November 10 and one paid account, which was used to launch more emails, on November 14. "In total, around 13,000 emails were sent to addresses on external recipient lists supplied by the sender," said Siozos. "The trial accounts sent 9,853 between November 7-9, and the paid account sent 3,723 on November 14. "We can’t tell how many people clicked, because the link was disguised to hide clicks. The recipient addresses were the sender’s own list." The CEO confirmed that LearnWorlds has since enhanced its bot checks and is placing holds on new accounts if their emails trigger fraud flags. "We have also examined the novel patterns the attackers used to create more advanced scoring mechanisms. We have reported the phishing website so it can be taken down. Further measures are currently being examined. "Anyone who signed in on the fake page should change their Google password and review their account’s recent security activity." The company's confirmation follows suspicions of foul play aired by users on social media. User reports included copies of the emails they received, which demanded urgent responses to fake copyright infringement claims related to YouTube content. Among the more prominent recipients of the copyright infringement email notices was Marcus Hutchins, of WannaCry fame, who claimed he received an email from the company’s official mail server. The email, a copy of which Hutchins posted to X, referenced general intellectual property and copyright infringements but failed to specify any specific issues. The email prompted the recipient to click a YouTube link to see the “supporting documentation submitted by LearnWorks.” The link predictably turned out to be a YouTube redirect link, taking recipients to a separate website that presents them with a real Google OAuth page. If the user completed the authorization flow, it would have handed attackers control of the target’s Google and YouTube accounts. “The site is designed to look like an official YouTube webpage that enables users to review copyright claims against their channel,” Hutchins wrote. “It contains both the YouTube and Google branding, as well as a very similar theme to the official YouTube one.” According to Hutchins' screenshots, and those posted by others, the attacker was actively tweaking the email style and format, although the underlying lure and authentication mechanisms remained the same. Hutchins didn't specify exactly how the phishing page steals access to targets’ Google accounts, although his description aligns with the steps often seen in OAuth consent phishing attacks. These involve malicious websites serving victims genuine OAuth pages, into which unwitting victims enter their credentials and pass the relevant MFA tests. Typical scenarios see permission pages appear immediately after, and if the user accepts, the malicious website is handed OAuth access tokens to the victim’s account, effectively the keys to control it. LearnWorlds, based in Cyprus, claims to have two million active monthly learners on its platform accessing 900,000 courses and more than 14,000 business customers. Since launching in 2014, it claims to have generated more than $1 billion in revenue. The business focuses on white-label courses - training programs developed by one business and sold to another, assuming the buyer will change the branding and present the course materials as their own. ®

US allies deepen deal to guard AI supply chains against 'hostile actors' (cough, China, cough)

16 November 2026 at 16:15
Nine countries including the UK have signed a US declaration pledging to secure supply chains for AI tech, as pressure grows for tighter export controls following the high-profile resignation of Nvidia CEO Jensen Huang. The US is firming up its Pax Silica initiative on AI and supply chain security. Under the San Francisco declaration, signed on November 15, the coalition says it is entering a new phase that turns shared principles into a working partnership with shared responsibilities. Pax Silica was announced in December last year, ostensibly for the participants to work together on vital parts of the global technology supply chain, including frontier foundation models, compute and semiconductors, advanced manufacturing, minerals refining and processing, and energy. But the list of participants was notable for not including one country – China – leading many to suspect that Pax Silica is yet another effort by Washington to squeeze out its main rival in the global technology and economic stakes. Hong-Kong based South China Morning Post refers to it as Trump's "campaign to China-proof AI and chip supply chains." The latest declaration states that "the security of our peoples cannot be separated from the security of our economies and those of our closest partners," and sets out what must be protected, as well as committing to "protect it together." It asserts that protection should cover key foundational technologies essential to economic security, listing "software, materials, knowledge, components and inputs…essential to AI supply chains." This includes anything that "theft, illicit transfer, or acquisition by hostile actors would confer a meaningful military or intelligence advantage, or undue economic leverage over the participants, or whose unauthorized transfer or misuse could pose significant risks to the economic and national security of the participants." Also listed is the protection of critical infrastructure such as ports, terminals, transport, communications, and logistics networks, and the ICT and energy infrastructure these depend upon. The declaration was jointly released by the US administration, and the governments of Australia, India, Israel, Japan, South Korea, Qatar, Singapore, United Arab Emirates, and the United Kingdom. "The know-how to design a chip, the machine that makes it, and the port that ships it can all be in different countries," the US under-secretary of state for economic affairs Jacob Helberg told the Financial Times. "What makes the San Francisco Declaration historic is that Pax Silica partners have agreed to treat the security of those technologies and the infrastructure behind them as a shared responsibility." This move comes amid reports that both Republican and Democrat lawmakers want to push through further measures to prevent China from accessing advanced chips for AI processing. It follows the recent news that Jensen Huang, the now former CEO of top GPU maker Nvidia, admitted to an affair with a woman suspected by the US of being a spy for Beijing. She worked for a Singapore-based company called Megaspeed that bought Nvidia hardware and was also suspected of making it available to Chinese users. US, Singaporean, and Malaysian police investigations are still probing whether chips were diverted or accessed remotely by Chinese firms. American efforts to control Chinese access to its technology started under the first Trump administration, and continued under President Joe Biden’s leadership. A report out last week claimed US export controls are not stopping AI labs in China from using restricted Nvidia chips, and many of their cutting-edge models were trained using Blackwell GPUs. It is understood that President Trump is to meet with China's premier Xi Jinping this week for bilateral talks on the sidelines of the Asia-Pacific Economic Co-operation (APEC) summit in Shenzhen. ®

❌