AI today, 5 November 2026: OpenAI agents' 948 Linux exploits, a data-centre backlash, an EU adviser on Gemini
11 things that happened in AI, each in under a minute. The day's news is about the cost and the limits of scaling AI. OpenAI's kernel run shows what automated offence costs, while politics, courts and geopolitics push back on data centres, press content and cross-border deals. Meanwhile labs are paying for content and chasing new compute deals.
-
Safety · Big
OpenAI says its agents built 948 Linux root exploits in one $1.2M run
OpenAI says a mid-October run of an autonomous pipeline on the Linux kernel produced working local privilege-escalation exploits (ones that turn a normal user into root) for 948 distinct bugs. API cost was about $1.21M (about $1,279 per bug), per Runtime Wire's summary, which says the pipeline also generated candidate patches for the 948 bugs. All figures are OpenAI-reported; patch status upstream is not confirmed here.Why it matters: It shows how cheaply AI can now mass-produce exploits, which helps defenders and attackers alike. Earlier context: PHP2Shell.
-
Policy · Big
AI data-centre stocks fall after Democrats win Congress; Ohio towns restrict new sites
After Democrats won the House and Senate on 3 November, data-centre-related stocks dropped on 4 November: Modine fell about 24% and Vertiv and GE Vernova about 11-12%, per Yahoo Finance. Goldman Sachs expects conditional moratoria from Democratic governors in states it says hold about 45% of planned US data-centre capacity. CNBC says CoreWeave and Iren fell about 4% in morning trading and links the move to Texas, where Gina Hinojosa won the governor's race on a promised data-centre moratorium. In Ohio, Fortune reports at least 14 municipalities passed data-centre bans or restrictions.Why it matters: Local politics, not chips, is becoming a constraint on the AI buildout. Anyone financing or siting compute should watch state-level rules.
Yahoo Finance · CNBC · Fortune
-
Policy
EU Court adviser: Gemini may infringe press publishers' rights in two ways
On 3 November Advocate General Maciej Szpunar issued a non-binding opinion in Case C-250/25 (Like Company v Google). He said a chatbot can infringe if it gets an article by bypassing a paywall or bot block, or if its answer copies the article's form so closely that readers no longer need the original. PPC Land reports facts stay free to use, and he said training is in principle covered by the text-and-data-mining exception. No judgment date is set.Why it matters: The Court usually follows its advisers, so this previews how Europe may treat AI answers built from news. Related: NYT v. Brave.
-
Business
Google to backstop about $19B of Anthropic's Kentucky data-centre leases
TeraWulf said on 4 November that Google will guarantee certain Anthropic lease obligations at a campus in Hawesville, Kentucky, supporting about $19B of contracted lease payments, per Unite.AI. The 20-year lease covers about 401 MW; the backstop ends once Anthropic holds investment-grade ratings for three months, and Anthropic must use Google-approved TPUs. Service starts in phases from late 2027.Why it matters: A top-rated company is lending its credit to help Anthropic fund compute, tying the two more closely. See the S-1 deep dive on its commitments.
-
Open source
Moonshot opens Kimi K3.1 Pro: 2.8T parameters, 1M-token context
Moonshot AI published Kimi K3.1 Pro on Hugging Face: a mixture-of-experts model with 2.8T total and about 104B active parameters, a roughly 1M-token context and text-plus-image input. Moonshot's own numbers include 94.2 on GPQA-Diamond and 65.7 on Terminal-Bench 4.0, but it trails GPT-6 Astra and Claude Opus 5.5 on HLE-Diamond (39.4 vs 65.8 and 60.9). It uses a custom licence, not a standard open-source one; all scores are vendor-reported.Why it matters: It is one of the largest weights anyone can download, but the custom licence and mixed benchmarks mean the headline claims need independent checks.
-
Business
China reportedly blocks SoftBank's $5.4B purchase of ABB's robotics unit
The Financial Times reported on 5 November, citing people familiar, that China's planning agency NDRC will not approve SoftBank's $5.4B deal for ABB Robotics, as relayed by investingLive. Grounds were not disclosed and SoftBank is exploring options; its shares fell about 5% in Tokyo. The EU had cleared the deal in March; China is the unit's largest market.Why it matters: It is a sign that physical-AI and robotics deals are now subject to geopolitical vetoes.
-
Products
OpenAI ships ChatGPT Images 2.6 with GIF creation and two new API models
OpenAI released ChatGPT Images 2.6 on 4 November, adding animated-GIF creation and reusable reference images. Per OpenAI's community post, the API gets GPT-Image-2.6 Flare (as little as six seconds per image, an OpenAI best-case figure) and Sunburst (higher quality and edit consistency).Why it matters: Image tools are becoming a mass-market feature, with faster models and reusable references aimed at everyday use.
-
Business
Anthropic signs its first paid news-licensing deal, with the Associated Press
Adweek reports Anthropic reached its first content licensing agreement with a digital news publisher, the AP, shortly before the midterms, letting Claude return live election updates credited to the AP. Terms and duration were not disclosed. Anthropic had argued that using publisher content is fair use, while OpenAI has more than a dozen such deals.Why it matters: It shows that real-time news access has a price even for labs that resisted paying, in the week an EU adviser questioned AI use of press content.
-
Research
Anthropic: Claude helped solve the first 3D structure of a vault particle from prokaryotic cells
Anthropic describes a lab project where Claude searched a database of about 15 billion predicted proteins, designed DNA constructs, cut purification from about three days to about 90 minutes and ran cryo-EM processing. The phage-derived vault has 96 copies of its main protein versus 78 in mammalian vaults, and its function is still unknown. Anthropic also notes Claude's mistakes, including a protocol that cost two weeks and a skipped processing step.Why it matters: A candid account of AI doing hands-on lab-adjacent science, including where humans had to rein it in.
-
Hardware
Qualcomm guides soft, but says data-centre revenue hit $650M and targets $5B next year
Per CNBC, Qualcomm reported $10.44B revenue (down 7%), handset chips down 26%, and data-centre revenue of $650M. It expects $5B of data-centre revenue in fiscal 2027; shares fell about 7% in extended trading on guidance below consensus. CEO Cristiano Amon said the company does not have a demand issue but a supply issue.Why it matters: Smartphone weakness is pushing Qualcomm toward AI inference chips, a bet it has yet to prove at scale.
-
Business
Apple reportedly cuts some engineers' daily AI spending cap from $300 to $200
Bloomberg reportedly said Apple lowered internal daily AI usage allowances for some engineers from $300 to $200 and is considering $100; limits could fall to $25 for some non-engineering teams, as summarised by The Neuron. Apple has not confirmed it; staff can request exceptions.Why it matters: Even the richest firms are rationing AI usage, a data point on the real cost of agentic work.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].