AI Weekly · Week 45, 2026 · 11.02 — 11.08AI Weekly

In shortOpenAI advanced on security and weights at once this week: GPT-6 Cyber is trained specifically to hunt vulnerabilities, able to find flaws, build exploit code and suggest patches, scoring 48.5% on ExploitGym against 32 for the previous GPT-5.6 Cyber. The autonomous pipeline found, verified and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. On the research side, Weight-Space Programming lets GPT-6 Astra read and generate raw neural-network weights directly without step-by-step mathematical optimisation; Astra identified from weights alone the four games and training language a network had learned, wrote weights from scratch for three simple games, and averaged 517 points on Breakout.

In shortAnthropic is still pursuing a pre-Thanksgiving IPO amid a tech-stock sell-off, planning to raise $80bn–$100bn at a valuation of $1.25tn–$1.75tn, below the $2tn the most optimistic investors hoped for; the earlier $1.5tn valuation had already drawn investor doubts and it is unclear whether the roadshow can start on schedule. On results, quarterly revenue was close to $17bn, twelve times a year earlier, but $17.5bn in operating costs exceeded revenue and net loss was $23.33bn against $12.1bn a year earlier, with the rise driven mainly by $5bn of locked-in short-term compute capacity and $2bn of employee donation matching.

In shortThe Democrats took both chambers in the midterms, with many campaigns centred on holding AI companies to account, and the industry now faces tighter regulation and slower data-centre construction. 43 of 44 local data-centre bans passed at the ballot, Texas governor-elect Gina Hinojosa said she plans to pause all data-centre approvals, and an NBC News poll found 76% of respondents held the related position. On capital, OpenAI completed a $11bn first funding round, the first tranche of its $1.25tn valuation round, which could grow to $20bn, below the original plan to raise $30bn at a $1.4tn valuation.

In shortAfter an 11-week investigation inside Anthropic, METR published a report saying its model improvement rate is about twice that of a year ago, with the past year's progress equal to about two historical years; researchers are confident the acceleration comes mainly from AI rather than simply more compute, but key evidence was heavily redacted. A separate audit found model capability growth has been about 1.9 times faster since October 2025, with AECI rising from about 15 to 29 points a year, and may have crossed its own acceleration warning threshold as early as 15 July 2026.

In shortSeveral vendors updated at once: Google DeepMind released Gemini Omni 1.2 Flash, extending video generation from 10 to 15 seconds and cutting the 720p price by 30% to $0.07 per second; Thinking Machines released Inkling 2 Small, a multimodal MoE model with 594B total and 17B active parameters; Mistral released Mistral Large 4 and expanded Mistral Studio; NVIDIA released Cosmos 3, four open-weight world-simulation models; and OpenAI released Images 2.6 and announced the acquisition of Pierre Computer Company to expand Codex tooling. On agents, the same open-weight model under different harnesses ranged from 30.0% to 81.1% pass rate on AppWorld, and Vals AI benchmarks show the cyberattack capability gap between open and closed models keeps narrowing.