IT News Roundup: Cold-Chain Cyberattack, Kimi K3.1 Pro Open Weights, 948 Linux Kernel Exploits — Week of Nov 02 – Nov 08, 2026

This week in IT was defined by a cold-chain cyberattack that crossed from the digital into the physical world, a second wave of supermarket refrigeration systems pushed into overdrive across ten European countries, and two of the year's largest financial thefts moving through courts and crypto chains. In AI, the open-weights race accelerated: Moonshot published the full 2.8-trillion-parameter Kimi K3.1 Pro, and OpenAI demonstrated a pipeline that produced 948 working Linux kernel exploits for roughly $1,300 each. Capital markets followed the story — Anthropic's IPO filing went public, an independent audit suggested the company's model progress is running nearly twice as fast as a year ago, and OpenAI closed an $11 billion round at a $1.25 trillion valuation.

Deep Dive: Europe's Cold-Chain Cyberattack Hits a Second Wave

On the night of 2 November, refrigeration and freezer systems at dozens of supermarkets across Belgium, Luxembourg, the Netherlands, Sweden, Estonia, Ireland, Portugal, Poland, Spain and Germany were pushed into uncontrolled heating, in what appears to be the second large-scale wave of the cold-chain cyberattack that began in France. Swedish chain ICA, among the affected stores, runs the same Danfoss controllers at the centre of the October incident, and German chains Edeka and CAP Markets also reported disruptions. As in the first wave, the attack forced refrigeration units into a continuous defrost cycle overnight, with temperatures rising far beyond safe levels and food stock spoiled.

The most serious incident of the week occurred at three blood banks of France's national blood agency, the Établissement français du sang (EFS). Attackers targeted the refrigeration units holding blood plasma for transfusion and attempted to heat them to 110 °C. Staff were able to intervene after the monitoring systems triggered, and the plasma was saved. EFS's refrigeration supplier, Coolworld, attributed the vulnerability to "a component supplied by a third party", extending the supply-chain dimension of the incident beyond the well-known Danfoss controllers.

The second wave follows an unprecedented first attack on 11 October, when up to 700 French supermarkets — including Intermarché, Netto, Carrefour, Monoprix, E.Leclerc and Coopérative U stores — were hit in a single night. The compromised refrigeration controllers forced the units into a defrost cycle with their contents still inside; alarms did not trigger, and many stores reported refrigerators climbing to 50 °C within six hours, with one store in Normandy reporting units reaching 80 °C. The heat melted equipment, started fires and set sprinklers running, and whole aisles of produce and frozen goods had to be discarded. Losses were estimated in the tens of millions of euros, and the October incident is widely regarded as the first cyberattack with a real-world effect on a national food supply chain.

The technical core of both waves is the same: web-management interfaces on the Danfoss refrigeration controllers are reachable directly from the public internet, and many installations were still using default credentials — "admin" as the username and the current date as the password. An attacker who logs in can trigger the defrost cycle remotely; because defrost is a normal operating state, no alarm is raised while the units run it continuously. In August, US authorities were investigating a similar heating incident at freezers in DeCA commissaries on military bases, suggesting the technique has been in use for months.

The incident is a stark reminder of how exposed operational technology remains. For IT professionals and homelab operators, the practical lessons are to inventory every OT and IoT device with a web interface and verify that none are reachable from the public internet; to search for default and date-based credentials across any device fleet; to segment control networks from the general LAN so a compromised management interface cannot reach field equipment; and to add monitoring for anomalous operating states — a refrigerator that enters defrost mode at 3 a.m. and stays there is an event worth an alert, not a log line.

No group has claimed responsibility for either wave, and the perpetrator remains unknown. The timing of the second wave — a month after the French incident — suggests the technique is being reused rather than improvised. Watch for vendor advisories from Danfoss and Coolworld on the affected controller firmware, any law-enforcement attribution, and whether the August US commissary incidents are formally linked to the European attacks. Sources: Reuters, The Guardian, Wikipedia.

AI Models & Weights

Kimi K3.1 Pro Open Weights Land on Hugging Face

Moonshot AI published the open weights for Kimi K3.1 Pro on Hugging Face on 4 November, weeks after the model's API launch. The release is a 2.8-trillion-parameter mixture-of-experts that activates 104 billion parameters per token, built on a 93-layer architecture with 896 experts (16 selected per token plus 2 shared) and a 1,048,576-token context window. Attention combines 69 Key-Distilled Attention layers with 24 Gated DeepSeek Sparse Attention layers, replacing the full-attention Gated MLA layers of the K3 predecessor to cut long-context cost. Weights are quantization-aware trained to MXFP4 with MXFP8 activations, and a 401-million-parameter MoonViT-V2 vision encoder makes the model natively multimodal. The release immediately topped Hugging Face's trending list.

Moonshot's self-reported benchmarks place K3.1 Pro among the strongest open-weight models for coding and agentic work: 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, 92.1 on BrowseComp, 70.0 on HLE-Full with tools and 94.2 on GPQA-Diamond, with independent evaluations still pending. The weights ship under a bespoke Kimi K3.1 Pro License rather than Apache 2.0, restricting certain commercial uses and derivative works — a continuation of the trend of restrictive licenses on large open-weight releases. API pricing on platform.kimi.ai is $2 per million input tokens and $10 per million output tokens, with a faster Highspeed tier at $4 and $20. Sources: Hugging Face, Kimi blog, Emergent.

For local inference, the practical news arrived two days later: llama.cpp build b12180 (6 November) added support for the Kimi-K3.1-Pro text model, implementing the DSA "lightning indexer" layers with a hybrid cache — a K-only MLA cache for the DSA layers and recurrent state for the KDA layers. Running the full 2.8T checkpoint requires a multi-node GPU setup, but the llama.cpp support means the GGUF ecosystem can follow for anyone who wants to experiment with the model's long-context and agentic behaviour on smaller hardware. Source: llama.cpp releases.

Inkling 2 Small Sets an Open-Weight ARC-AGI Record

ARC Prize published results for Thinking Machines' Inkling 2 Small on 6 November: 96.5% on ARC-AGI-1 semi-private and 81.7% on ARC-AGI-2 semi-private at XHigh reasoning, exceeding every open-weight model the prize has evaluated to date and setting a new state of the art in that category. Thinking Machines says the model's weights are pending release. Source: ARC Prize.

Alongside the benchmark results, the company described how it caught early training checkpoints that had learned to bypass the model's guardrails — a useful data point in the effort to catch safety regressions before models ship rather than after. Source: LLM Stats.

P-JiT: A $21,038 Text-to-Image Model

Linum released P-JiT on 4 November, a 2.18-billion-parameter text-to-image model that works directly in pixel space. The full pretraining and post-training run cost $21,038 on 8,415 H100 hours, and both weights and code are Apache 2.0. At that size the model is practical on consumer GPUs, and the published training cost makes it one of the cheapest complete image-generation recipes of the year. Source: AI/TLDR.

The Rest of the Week

OpenAI's Cyber Model Produces 948 Linux Kernel Privilege-Escalation Exploits

OpenAI published "A Thousand Fewer Paths to Root" and the GPT-6 Cyber system card on 4–5 November, describing an automated vulnerability-discovery pipeline run on an upcoming cyber model. In a roughly 30-hour run in mid-October, the pipeline analysed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 as real vulnerabilities and produced 948 working local privilege-escalation exploits that let an unprivileged local user gain root. The run consumed about $1.2 million in API credits — roughly $1,279 per distinct flaw — and OpenAI generated candidate patches for each exploit for maintainer review. Sources: OpenAI, Phoronix.

Two details drew particular attention. First, OpenAI's GPT-5.5 Cyber model found more than five times more novel LPEs than the earlier GPT-5.6 Sol model even though the kernel had been hardened and patched for three months between scans — a sign that discovery capability is compounding faster than the patching process. Second, OpenAI's researchers argued that within a few years frontier models will likely be able to find existing vulnerabilities on their own, and that less capable models may be able to develop exploits once they are disclosed.

For Linux operators — homelab hosts included — the practical takeaway is that local privilege escalation is becoming a much faster-moving threat class. Keeping kernels and userspace patched promptly, restricting local access for untrusted users and containers, and treating every kernel release as security-critical are the immediate mitigations while the ecosystem digests the patch volume.

Debian and Ubuntu Patch an APT Flaw That Lets MitM Attackers Run Code as Root

Debian, Ubuntu and APT's developers published advisories on 4 November for CVE-2026-109518, a flaw in APT that lets a man-in-the-middle attacker swap a verified package for a malicious one and execute arbitrary code as root. The attack works when APT retries a failed download — the default behaviour from Debian 12 and Ubuntu 22.04 onward — using HTTP redirects and ordinary transient server errors to make the retry fetch attacker-controlled content.

Ubuntu rates the flaw high priority with a CVSS 4.0 score of 7.7 and published new APT builds for Ubuntu 22.04, 24.04, 26.04 and 26.10 on the same day; Debian shipped fixes including a 2.2.5 build for Debian 11, whose long-term support ended in August. The vector is anyone who can alter HTTP traffic between APT and the package server — a malicious mirror or proxy. No exploitation has been observed, and the advisories describe the attack over HTTP only, leaving open whether HTTPS-served sources are affected. Sources: The Hacker News, Phoronix.

The flaw was reported by Max Justicz. For homelab and server operators the fix is a simple APT upgrade, but it is worth verifying that mirrors and proxies in the update path are trusted, since the entire attack model depends on controlling the download channel. Source: Debian security.

Android's November Bulletin Ships 216 Fixes, Headlined by a Zero-Click RCE

Google's November 2026 Android security bulletin, published on 3 November, ships 216 fixes, 23 of them rated critical. The headline is CVE-2026-58868, a critical remote code execution in the Android Framework: a heap buffer overflow leading to an out-of-bounds write in the onOutputReceived function of FrameDecoder.cpp while processing media frames. The flaw requires no user interaction and no additional privileges, and affects Android 14 through 17. A companion critical flaw, CVE-2026-58882, is a denial of service in SMS WAP Push handling. Sources: SecurityWeek, HIFY Radar.

The fix ships with security patch level 2026-11-01 or later (covering 119 flaws in the runtime, framework and system), with 2026-11-05 recommended to cover the full bulletin, including 97 kernel and hardware fixes from Arm, Imagination Technologies, MediaTek, Unisoc and Qualcomm. Seventeen flaws had already shipped early via Google Play system updates. Google reported no known attacks as of the bulletin, but its data shows over 40% of Android devices run versions too old to receive the fix at all.

For enterprises and BYOD fleets, the action item is to check each device's security patch level — Settings, About phone, Android security update — and require at least 2026-11-01, targeting 2026-11-05. Google also published companion bulletins for Android XR and Wear OS on Monday.

Crédit Mutuel Arkéa Confirms a €100 Million Cyber Heist

French cooperative bank Crédit Mutuel Arkéa confirmed on 5 November that hackers stole around €100 million ($112.6 million) in a cyber heist, a rare case of direct theft from a bank's own funds. The incident first surfaced on 2 November, when researcher ZachXBT reported a suspected theft of more than $106 million from an unnamed EU entity, later raising his estimate to over $133 million; he named the bank late on 3 November, and the bank confirmed the attack the following morning. French prosecutors opened an investigation on 5 November. Sources: Reuters, Halborn.

According to reporting by French journalist Grégory Raymond, the attackers moved money from the bank's own account to accounts they had opened at centralized exchanges under fake identities; on-chain tracing shows the fiat was converted into USDC and withdrawn to external wallets. The Brittany-based bank has about €211.5 billion in assets and more than 5.8 million members and customers, and said customer funds were not affected.

The case is notable for its size and its mechanics: the money left through the bank's own transfer channels, which points to compromised internal authorisations rather than a perimeter breach. Watch for the outcome of the French investigation and whether the USDC proceeds can be frozen before further mixing.

North Korea Denies Role in the $6.3 Billion Binance Heist

North Korea denied on 2–3 November any involvement in the $6.3 billion Binance heist, dismissing the US allegations as "absurd rhetoric" and a "grave political provocation" in a statement carried by state agency KCNA and attributed to Kim Yo Jong. The denial came after the FBI attributed the theft — the largest cryptocurrency heist on record, surpassing the $1.5 billion Bybit hack of 2025 — to TraderTraitor, a North Korea-linked group, on 1 November. Sources: Reuters, Wikipedia.

Binance put the impact at about $6.34 billion, with Ethereum and Bitcoin the most affected blockchains. Co-CEO Richard Teng said on 3 November that the attack began with a fake job interview: an engineer downloaded malicious software — including a tampered Homebrew package — during the process. Binance also described a poisoned coding task and a hijacked internal AI tool used to reach the exchange's wallet systems, with preparations dating back to June. Binance said outflows have normalised and reported that nearly $3 billion of the stolen assets were frozen. Source: CNBC.

For security teams, the heist is a masterclass in how social engineering is now paired with supply-chain compromise: initial access was a job interview, the payload was a package manager, and lateral movement ran through an internal AI tool. Screening of hiring and contractor pipelines, strict controls on what development machines can execute, and isolation of internal AI tooling from financial systems are the direct lessons.

Anthropic's IPO Filing Meets a METR Audit and Investor Doubts

Anthropic's IPO filing went public on Monday, 2 November, with reports pointing to a potential valuation of up to $2 trillion and a debut targeted for mid-November. The filing drew fresh scrutiny on Thursday when METR, the Berkeley-based nonprofit that evaluates frontier models, published an independent assessment of AI-driven research acceleration at Anthropic based on an 11-week review (12 August to 29 October) with access to confidential internal data. Source: Reuters.

METR's central finding: as of October 2026, Anthropic is likely improving its models' capabilities on well-specified tasks at about 1.9 times the rate achieved between March 2024 and October 2025, with a 95% interval of 1.5 to 2.8x. METR judged that AI use likely explains most of the speedup — perhaps around 80% — and found convincing evidence that faster growth in training compute did not cause it. The report also noted "signs of life" for fully autonomous AI contributions to research that are not yet large, and saw no significant algorithmic, compute or data bottleneck over the next 12 months. Source: Baozilla's AI timeline.

The report's reception was complicated by process: Anthropic held unilateral redaction rights and removed a majority of the evidence METR found most informative, including per-stage acceleration estimates and the specific benchmark names; a planned audit of compute allocation was reportedly cancelled mid-pilot. Co-founder Sam McCandlish said the redactions cover commercially sensitive material, and Anthropic promised its own analysis in the coming weeks. Potential investors have separately raised doubts about the IPO's timing, and prediction-market odds of the listing by November 30 fell from 84% to 60% on Thursday. Source: Yahoo Finance.

Week in Numbers

• OpenAI closed an $11 billion funding round at a $1.25 trillion valuation on 6 November — the first tranche of a round initially targeting $30 billion or more at $1.4 trillion. Reuters
• Google acquired Crusoe's AI factory site in southeast Wyoming for a reported ~$17 billion on 6 November, after the neocloud struggled to lock in customers for the project. TipRanks
• 948 Linux kernel local privilege-escalation exploits were produced for $1.2 million in API credits — about $1,279 per distinct flaw. OpenAI
• Android's November bulletin shipped 216 fixes, 23 of them critical, while over 40% of Android devices run versions too old to receive the patch. HIFY Radar
• Wordfence observed more than 18,500 targeting requests against the WordPress core TOCTOU flaw (CVE-2026-119977, added to CISA's KEV catalog on 4 November) in the first 21 hours after disclosure. Cyber Recaps