AI News Top 10
5 November 2026
OpenAI says an agentic pipeline found, reproduced and exploited 948 Linux kernel bugs in a single 30-hour run that cost USD 1.2 million.
OpenAI pipeline finds and exploits 948 Linux kernel bugs in one 30-hour run
OpenAI detailed an agentic pipeline that analysed 2,308 candidate findings in the Linux kernel with an upcoming cyber model, reproduced 1,724 of them dynamically and produced 948 local privilege escalation exploits that let a normal user gain root. The run took 30 hours and cost USD 1.2 million in API usage. It builds on the 24 LPEs GPT-5.5-Cyber found in June 2026.
MIT: GPT-6 Astra reads and rewrites the weights of other models
MIT researchers report that GPT-6 Astra can read and rewrite the internal weights of other networks without conventional training. Astra identified from raw weights which of four games a network played, which language a small model learned and which of two models was fine-tuned to flatter users. Weights generated by Astra beat trained models 517 to 25 on Breakout; on Qwen3.5-0.8B a change to 0.016% of the weights lifted MATH-500 from 7.9% to 48.1% and instruction following from 21% to 62%, while a third of answers still got stuck in reasoning loops.
OpenAI policy chief frames 2028 as 'The AI Election'
OpenAI chief global affairs officer Chris Lehane published an essay that compares AI's economic impact to Roosevelt's New Deal and calls on politicians to find a 'third way' so people share in the gains. He lists five priorities: worker protection and education, AI safety, data centre development, child safety and tax incentives for AI skills. The essay appeared directly after midterms in which Democrats took Congress and data centre opposition proved bipartisan; the AI Now Institute calls the New Deal framing ahistorical at best.
OpenAI renames data centres 'AI compute campuses' after midterms
OpenAI is preparing a post-midterm campaign that renames data centres 'AI compute campuses', Axios reported, combined with lobbying and image repair. A Gallup poll from July found that up to 70% of US adults oppose construction in their own area, across party lines. Lehane calls the debate healthy and says communities must genuinely benefit; critics describe the mix of rebranding and self-proposed safety rules as regulatory capture.
Anthropic carries USD 895 billion in future compute commitments
Anthropic reported about USD 895.4 billion in purchase and lease obligations, almost entirely compute, running to 2045 in an amended IPO prospectus from 2 November; USD 214 billion of that is cancellable under conditions and some contracts must be paid even when unused. Alphabet posted negative free cash flow for the second quarter in a row after USD 51.4 billion of capex in Q3. Among buyers, 56% saw a supplier charge an AI surcharge or move existing features to a pricier tier, and 38% delayed or cancelled an AI project over cost overruns.
Parts of the EU AI Act now apply in Northern Ireland
The UK government confirmed that EU rules for AI apply in Northern Ireland, but only for physical goods with an embedded AI system that are placed on the NI market, a consequence of the Windsor Framework that keeps Northern Ireland in the EU single market for goods. Online or downloadable AI tools fall outside it, and the obligations do not touch the developers of such tools. Cabinet minister Hamish Falconer says no new regulatory border with the rest of the UK arises.
Yale: 36% of voters asked a chatbot how to vote, and the models differ sharply
Almost 2,500 voters in 35 states with Senate races questioned ChatGPT, Claude, Grok or Gemini about the elections a few weeks before the midterms. A third asked a model for help choosing a candidate. Once users shared preferences, ChatGPT recommended a candidate in 79% of conversations and Claude in 40%; without any user information only Grok did so, in 14% of conversations.
Ricursive releases Ripple-65B Preview under Apache 2.0
Ricursive Intelligence published Ripple-65B Preview, a dense open-weight language model under Apache 2.0. It scores 80.5 on Terminal-Bench 2.1 and, according to its maker, leads US open models up to fifteen times its size. VentureBeat reports the model also claims the US open-source lead on two large coding benchmarks.
Waymo leaves the Uber app, hours before Uber says it alone can make robotaxis pay
Waymo stopped offering bookable robotaxis through Uber in Austin and Atlanta; co-CEO Tekedra Mawakana told Bloomberg that Uber users who ordered a robotaxi were not guaranteed a Waymo. On the same day Uber CEO Dara Khosrowshahi presented quarterly figures with the claim that Uber's platform gets 30% more rides per vehicle per day than apps of autonomous providers. Uber wants about 2,000 robotaxis on the road in 2027 and a large expansion from 2028.
ChatGPT now makes animated GIFs, and OpenAI processes 4 billion images a week
OpenAI is rolling out Images 2.6 to all ChatGPT and Codex users on web, desktop and mobile. ChatGPT can now make a GIF from a prompt or an uploaded photo in one step, with adjustable speed, frame-by-frame editing and direct sharing. OpenAI says users generate more than 4 billion images per week through ChatGPT and the API, up from 3 billion in September; GPT-Image-2.6 Flare and Sunburst arrive in the API.
Key Trends
- • Agents move from finding bugs to exploiting them: one pipeline produced 948 kernel exploits, which makes patch speed the bottleneck.
- • Model weights become an interface: Astra edits other networks directly, so changing a model's behaviour no longer needs a training run.
- • Compute commitments now dominate balance sheets: Anthropic carries USD 895 billion to 2045 while Alphabet books negative free cash flow.
- • OpenAI runs two political tracks at once: an election essay and a rebrand of data centres as AI compute campuses.
- • EU rules travel through the Windsor Framework: makers of smart hardware need one EU-compliant variant for Northern Ireland as well.
Wat vond je van dit bericht?
