This brief covers AI news from 2026-10-29 UTC.
Google’s new frontier model is aimed at finding and fixing vulnerabilities, and it is not open to everyone yet.
Google ships Gemini 4 Carbon, first to cyber defenders
What happened: Google DeepMind announced Gemini 4 Carbon, which ranks first on Artificial Analysis Intelligence Index v5.0 at 59 and leads SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%. Access starts with a subset of cyber defenders through the Fairwind Program before wider release.
Why it matters: Security teams get a vulnerability-finding model before general developers do, so plan patch and audit tooling around gated access rather than immediate API availability.
Kimi K3.1 Pro opens a 1M-token window to developers
What happened: Moonshot’s Kimi released K3.1 Pro, a 3T-class open-weight model with a 1-million-token context, available on Kimi.ai, Kimi Code and the Kimi API, with full weights promised by Nov 4. It scores 65.7% on Terminal-Bench 4.0, behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%.
Why it matters: A million-token open-weight model with near-frontier terminal scores gives teams a self-hostable option for long agent runs once weights land next week.
GitHub Universe recasts the platform around agent fleets
What happened: At GitHub Universe, CEO Jay Parikh described GitHub as an end-to-end software delivery system where agents implement and verify changes while developers set direction, and introduced products including the GitHub Copilot app.
Why it matters: If your CI and review flow assumes humans push commits, GitHub is rebuilding the primitives underneath it for agents that ship completed work.
Android Studio lets Claude and Codex into the IDE
What happened: Google’s Android Studio Rabbit 2 adds a Bring Your Own Agent feature using the open Agent Client Protocol, letting developers run more than 40 coding agents including Claude, Codex and Google Antigravity with access to project structure, the emulator and Logcat.
Why it matters: Android developers can keep their preferred coding agent and still reach IDE internals, which lowers the cost of switching tools mid-project.
Microsoft booked $10B from OpenAI in one quarter
What happened: Microsoft said in a regulatory filing that it generated $10 billion in revenue from commercial arrangements with OpenAI, including revenue sharing, in the quarter ended September 30, nearly 11% of its $92.3 billion quarterly revenue. OpenAI-related revenue hit $24.1 billion for fiscal 2026.
Why it matters: The revenue split shows how much of Microsoft’s AI business still runs through one partner, which shapes Azure pricing and capacity for anyone building on it.
InfraDev benchmark grades agents on real vLLM and Megatron PRs
What happened: Researchers released InfraDev, a benchmark of 36 tasks built from real pull requests to Megatron-LM, vLLM and verl, where agents must reimplement multi-file changes that run correctly across many GPUs. Grading uses hidden probes plus a prompt stating every interface the probes call.
Why it matters: It gives teams a way to test whether a coding agent can actually touch training and serving infrastructure, not just single kernels.
Third Circuit rules Westlaw headnotes are not fair use for AI
What happened: The Third Circuit affirmed partial summary judgment in Thomson Reuters v. ROSS Intelligence, holding Westlaw’s editorial headnotes are copyrightable and that ROSS’s use of them to train its legal research tool was not fair use. The court said judicial opinions themselves remain free to use.
Why it matters: Training data built from another company’s editorial annotations carries appellate-level risk, so audit datasets that layer commentary over public records.