3 November 2026Critical Linux TCP flaw lets remote attackers run code; patch now
011 min read
Critical Linux TCP flaw lets remote attackers run code; patch now
Red Hat and Canonical have disclosed fixes for CVE-2026-98432, a use-after-free bug in the Linux kernel’s IPv4 TCP handling that a remote, unauthenticated attacker can use to execute code. The kernel CVE team rates it 9.8, and Red Hat has moved its severity to critical. The flaw has existed since Linux 4.4. Red Hat has published fixed kernels for RHEL 8, 9 and 10 under specific errata; installing them requires a reboot. Ubuntu’s fixes for 22.04, 24.04 and 26.04 are only in the opt-in -proposed testing channel, so they will not arrive automatically and Canonical has not given a regular-release date. Administrators should install the fixed kernel, reboot, and verify the running version with uname -r. Because any listening IPv4 TCP port may be enough, this is a high-priority patch for internet-facing Linux estates.
021 min read
Synology warns NAS ransomware attacks exploit already-patched DSM flaw
Synology is telling customers to update DiskStation Manager after ransomware attacks hit NAS devices running outdated DSM releases. The company says attackers are exploiting a vulnerability that was fixed in December 2024; although it did not identify the exact flaw, it linked to an advisory covering CVE-2024-10441, a 9.8-rated remote-code-execution issue in the system plugin daemon, plus two lower-rated bugs. Customers report C3RB3R ransomware encrypting files with a .L0CK3D extension. The practical exposure is any internet-reachable Synology NAS left on an old DSM build. Synology lists fixed versions: DSM 7.2.2-72806-1 or later, 7.2.1-69057-6 or later, 7.2-64570-4 or later, 7.1.1-42962-7 or later, and 6.2.4-25556-8 or later. Administrators should upgrade to the matching fixed build and check that backups are isolated.
031 min read
Refrigeration cyberattacks force European supermarkets to pull food
Attackers remotely altered refrigeration settings at supermarkets in Belgium, Sweden, Estonia and other European countries overnight into 2 November, warming chillers and freezers. No chain reported more than several dozen affected stores, but some retailers discarded or withdrew chilled and frozen food. Danfoss, whose controls were involved, also had products in a French incident three weeks earlier that hit hundreds of supermarkets; no one had claimed the new attacks and a link was not established. The affected equipment is internet-connected refrigeration control gear that contractors use for remote monitoring and repair. In Belgium, units were switched to defrost mode; in Sweden and Estonia, setpoints were raised to 50°C, though one Estonian store recovered quickly without product loss. Danfoss had posted no advisory by 3 November; CERT-FR’s October alert still offered the most detailed operator steps. Retailers should treat refrigeration OT as exposed critical infrastructure and use that alert as guidance.
041 min read
NCSC warns AI-powered ransomware against UK critical infrastructure by 2028
The UK’s National Cyber Security Centre has published a State of the Cyber Nation report warning that critical national infrastructure should prepare for a ransomware attack involving AI-powered malware by 2028. The NCSC assesses that such use is highly likely, and its cyber chief says years of choosing cheap technology over secure technology have left organisations exposed. The report calls for significant changes and urgent investment in resilience rather than treating AI as a future-only concern. For business leaders, the message is that AI-enabled attackers may find and exploit longstanding weaknesses at greater speed and scale, so exposure management, patching and incident readiness need board attention. The warning is a forward-looking assessment, not evidence of a specific attack. Firms in and around critical infrastructure should map dependencies, close known gaps and test response plans; all organisations should assume AI will sharpen existing threats.
2 November 2026GitLab RCE flaw exploited; over 1,000 servers still unpatched
011 min read
GitLab RCE flaw exploited; over 1,000 servers still unpatched
GitLab has disclosed a maximum-severity vulnerability, CVE-2026-97229, in the CI/CD regular-expression compiler. The flaw is an integer overflow that can let an unauthenticated attacker execute arbitrary code through a specially crafted regex in a pipeline configuration. Patches shipped on October 12 in GitLab Community and Enterprise Edition 19.4.2, 19.3.5 and 19.2.8. CISA has confirmed in-the-wild exploitation and added the bug to its Known Exploited Vulnerabilities catalog, requiring federal agencies to secure servers by Sunday. Shadowserver still sees more than 1,000 exposed, unpatched GitLab servers. GitLab.com is already patched, and GitLab Dedicated customers need not act. Self-managed administrators should upgrade immediately to one of the fixed releases. Because CI/CD systems hold source code, credentials and deployment rights, a compromise can spread quickly; restrict internet exposure and monitor for suspicious pipeline activity while patching.
021 min read
Ivanti patches two Connect Secure zero-days used in attacks
Ivanti has issued security updates for two Connect Secure remote-access gateway flaws that were used in zero-day attacks. One, CVE-2026-111357, is a high-severity privilege-escalation bug; the other, CVE-2026-118527, lets authenticated attackers upload files to an unpatched device. According to Volexity, a threat actor tracked as UTA0576 chained these with an authentication bypass, CVE-2026-104935, patched last month, to obtain unauthenticated root-level code execution on fully patched appliances from at least October 3. The actor reportedly used no malware or web shell, re-exploited the device for each command and tampered with logs to evade Ivanti's Integrity Checker Tool. Defused confirmed the findings. Organizations running Ivanti Connect Secure should install the new updates immediately, review whether their gateways were exposed, inspect logs for tampering and tighten access until they have confirmed remediation. These gateways sit at the network edge and often carry privileged access, so delay is risky.
031 min read
OpenAI to merge ChatGPT Chat and Work tabs and pool tokens
OpenAI is preparing to fold ChatGPT's separate Chat and Work tabs into one interface, starting with selected users before a broader rollout, according to an email sent to customers. The company says the change will let people move from exploring an idea to having ChatGPT carry it out in the same place, bringing in files and plugins and asking the assistant to build a website, create a presentation or perform web tasks while the user reviews the work. Codex will remain a separate desktop experience. OpenAI also said ChatGPT and Codex will share one token allowance pool; most users should not notice, but depending on usage and model, some may need to wait longer for resets. For businesses, the shift could simplify adoption and workflow design, but it also makes usage governance more important because heavy Codex use may affect ChatGPT capacity.
041 min read
Google disrupts 1.3M-device Android botnet; TV box backdoor remains
Google has dismantled the command-and-control servers of an Android botnet spanning more than 1.3 million devices, mostly TV set-top boxes. Google says it is highly confident that manufacturer Videostrong deliberately built the x.cpe backdoor into many models and brands and sold access to criminals, who used the boxes for ad fraud and to route other people's traffic through owners' home connections. The backdoor ran as a system app, weakened Play Protect checks, allowed silent app installs, remote commands, screen capture and reverse shells. Google's legal action removed control domains for the botnet and more than a dozen malware families and may cut off access to 10 million more infected devices. Google banned Videostrong from certified Android devices and software supply. The backdoor remains on affected boxes and owners cannot remove it, so businesses using such devices need to account for that risk and review supply-chain trust.
051 min read
Anthropic limits Claude Code recreations of existing games
Anthropic has changed Claude Code's system prompt on Opus 5.5 with two sentences intended to stop the tool from recreating specific existing video games. The guidance tells Claude to build original games, games within a genre, and games based on the user's own designs and assets, but not to reproduce a particular title even if it is renamed or presented as original. Anthropic engineer Addy Osmani said the policy reflects the company's rules against intellectual-property infringement and that legitimate original work should continue to function. Reports nonetheless describe some developers being blocked on their own projects, including anti-cheat and reverse-engineering work. The move arrives amid a surge of AI-assisted game decompilation and porting. For businesses, it signals that AI coding tools will increasingly enforce IP boundaries, so teams should document ownership and expect occasional false refusals when using Claude Code for game-related work.
1 November 2026GitLab CI/CD flaw: patch to 19.3.5 after attempted RCE
011 min read
GitLab CI/CD flaw: patch to 19.3.5 after attempted RCE
GitLab administrators should treat version 19.3.4 as unsafe and move to 19.3.5, the fixed release for CVE-2026-97229. The maximum-severity bug is an integer overflow in the CI/CD regular-expression compiler Onigmo, bundled with Ruby; unauthenticated attackers can run arbitrary code by submitting a crafted regex in a pipeline configuration. Alpine Linux, whose GitLab server hosts development repositories and build recipes, faced repeated exploitation attempts on October 20–21. It found no proof that code ran or that packages were altered, but missing logs meant it could not exclude a compromise, so it rebuilt the instance and reset all passwords. The lesson for businesses is that developer infrastructure is a supply-chain asset: patch quickly, keep outbound and execution logs, and be ready to rebuild rather than assume absence of evidence is safety.
021 min read
Reports: OpenAI cuts new funding target to $20B
Unconfirmed press reports say OpenAI has reduced the target for its latest fundraising to $20 billion, down from earlier reported goals of $30 billion, and that the first close may be smaller still. The Financial Times reported a group of investors from the United Arab Emirates leading the round, while The Wall Street Journal said about $11 billion had been secured and some large investors responded coolly. A later retail offering for wealthy individuals could lift the total toward the target. OpenAI did not comment in the report. For business buyers, the signal is not just about one vendor’s balance sheet: it points to tighter scrutiny of AI infrastructure spending and could affect pricing, capacity commitments and IPO timing across the AI supply chain. Treat the figures as reported, not confirmed.
031 min read
Northflank breach: rotate all secrets and AI API keys
Northflank has told customers to assume every environment variable stored on its cloud application platform was decrypted and to disable or rotate all keys immediately. The warning followed a breach in which multiple customers reported unauthorised use of AI API keys held on the platform. The critical exposure is credentials such as database passwords, cloud tokens and AI provider keys, which can enable further intrusions or unexpected charges. Affected organisations should rotate secrets now, review AI-provider usage and billing records for anomalies, and report suspicious activity to Northflank and the relevant AI vendors. The incident shows how a hosting provider can become a credential choke point for AI-dependent operations, so buyers should ask vendors about secret isolation and incident notification.
041 min read
Study: multi-agent trading systems can be poisoned
Research on LLM-based multi-agent trading systems finds that a single corrupted signal can travel from open-web news or social media to the final trade, causing financial loss. The study tests two adversaries that need no model internals: one plants manipulated content in web sources, the other tampers with an agent’s prompt. Across four roles—Analyst, Researcher, Trader and Risk Manager—and four communication topologies, no architecture was inherently robust. Majority voting helped only conditionally, depending on agreement among benign agents, and compromising a terminal decision node could bypass upstream safeguards. For financial firms moving these systems from prototypes to live assets, the practical message is to treat every external input and every agent prompt as attack surface, test role-specific and topology-specific failure modes, and avoid assuming that layered agents or voting alone provide safety.
051 min read
Power caps may not cut LLM serving energy, study finds
A study of LLM serving on an NVIDIA H200 under vLLM finds that GPU power caps do little for autoregressive decode, the phase that dominates serving. Decode kernels wait on HBM and the GPU idles between launches, so it draws less than half of the 700W rating and throughput changes only slightly under the caps tested. The cap engages mainly when a large batch runs a long generation. Locking the SM clock instead lowered power without slowing memory-bound decode and recovered up to 32% of decode energy with minimal throughput loss. The authors identify three architecture-dependent frequency-scaling classes and report that recurrent architectures pay no prefill penalty once prefix caching is excluded. For AI operators, the result suggests energy-saving plans should be phase-aware and not rely on power caps alone; clock policy and model architecture can matter more for cost and efficiency.