AI progress — 5 November 2026

OpenAI says an autonomous pipeline found, validated and patched 948 local privilege-escalation vulnerabilities in the Linux kernel. If the claim holds up, vulnerability discovery and patch generation at that volume is now something a single lab can run end to end. Elsewhere, US election results are pushing data-center siting and AI regulation into politics, and llama.cpp shipped a sizeable release.

New models and papers

• A thousand fewer paths to root — OpenAI describes an autonomous pipeline that found, validated and generated patches for 948 local privilege-escalation bugs in the Linux kernel. Why it matters: kernel maintainers and security teams should expect machine-generated reports and patches at this scale; the figures are OpenAI’s own.
• SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding — a paper on training parallel drafters for speculative decoding with on-policy distillation, aimed at longer acceptance lengths. Why it matters: speculative decoding speed-ups depend on acceptance length, so better drafters would cut inference cost without changing the target model; results are the authors’ to report.
• Progress on provable inference for large language models — an Anthropic research post on provable inference for LLMs (the candidate feed gives only the title). Why it matters: verifying that a provider ran the model it claims is a gap for anyone buying inference; read the post for what is actually demonstrated.

New tools and software

• llama.cpp v0.9.0 — adds OCP FP8 E4M3 quantization, a flash_attn_ext_rows op for multi-sequence KV-cache inference, four new architectures including Mistral Large 4, and prefill optimizations for Gemma 4 E2B/E4B. Why it matters: local serving of several concurrent sequences gets more efficient, and more new models run without conversion workarounds.
• LangGraph 1.2.20 — each interrupt() in a task now gets its own id, and a strict resume mode refuses stale interrupt answers. Why it matters: human-in-the-loop agents with several pending approvals can no longer resume with an answer meant for an older prompt.
• GitHub Copilot in Xcode’s coding assistant — Copilot is available inside Xcode’s built-in assistant, in public preview. Why it matters: Xcode users can use Copilot without leaving the native workflow.

Overall digest

• Texas data-center buildout may stall — MIT Technology Review reports that Democratic wins in the Texas Senate and governor races may jeopardize the state’s position as the fastest-growing US data-center market. Why it matters: teams planning capacity in Texas face policy risk that was not priced in a week ago.
• Big Tech disputes the AI backlash reading of the vote — Ars Technica says the industry is framing voter support for restricting AI and data centers as anti-Trump rather than anti-tech. Why it matters: this is the industry’s framing, and it will shape how hard companies lobby against new restrictions.
• Apple reportedly cuts internal AI budgets — TechCrunch reports Apple is significantly reducing spending on AI for its engineering teams; the report is unconfirmed. Why it matters: it is a data point on how a large company prices AI tooling for its own developers, nothing more yet.