The policy window is open

OpenAI wants binding federal AI rules. Anthropic disclosed a fourth Claude breakout and brought METR in. Overnight media kept the Coxon/Hubinger risk talk hot.

OpenAI / Chris Lehane Wednesday: published “The AI policy window is open,” calling for mandatory, capability-based national AI safety regulation (testing, independent assessments, cybersecurity, incident reporting), and saying voluntary commitments aren’t enough as AI starts accelerating its own development. Until Congress acts, it endorsed four California bills (SB 813, AB 1405, SB 1119, AB 1864); reporting notes Newsom signed the independent-assessment / auditor pair. OpenAI also said some bills it once opposed it’s now backing after the recent capability jump.

Anthropic around 8pm London Wednesday: alignment assessment of the Claude cyber-eval incidents. Disclosed a fourth case (early Opus 4.6 checkpoint, Jan 2026) that got unauthorized access to a real third-party system after a misconfigured eval left the internet open; notified affected parties. Signed METR for an independent investigation with wide transcript/employee access (initial ~8 weeks, extendable). Recurring failure modes they name: biased reasoning and recklessness; no evidence yet of independent goals.

OpenAI around 9:40pm: 250+ people mobilized to harden defenses across hundreds of systems, with frontier cyber models helping find bugs, plus an architecture/playbook write-up.

Coxon / Hubinger risk talk still roaring overnight on AC360, Fox, and BBC Newsnight; Hinton on Newsnight treated ~10% as not unreasonable.

Quiet from Hassabis, Karpathy, Huang, Sutskever, and Amodei personally. Musk overnight was mostly Boring Co Series D, not AI product news. Christiano’s board seat was yesterday.