METR Proposes Framework for Third-Party Evidence on Frontier Models' Loss-of-Control Risks
BethMayBarnes · x · 2026-10-15
METR has published its vision for working with frontier AI developers to openly assess and share evidence about loss-of-control risks in unreleased models. Asks: transparency on scope/redactions/limitations, staff freedom to share in-scope info, and rapid publication, with companies controlling redactions. Beth Barnes stresses it's voluntary, offers no 'assurance of safety,' and argues pre-deployment evaluation is inferior to addressing risks while models are still unreleased. She also urges companies to work with multiple third parties and favor full public release of logs over third-party involvement.
Related event: METR Proposes Framework for Evaluating Frontier AI Loss-of-Control Risks(2 posts)→
More from Safety
- After FBI Arrest of OpenAI Engineer, Grok-Powered OSINT Dig Surfaces His Saudi Ties — ctjlewis · 2026-10-15
- "No accountable human, no deployment": Chesterman on who's liable when AI agents go rogue — ProfChesterman · 2026-10-15
- Gemini 4 Argon survives 83% of coding prompt injection attacks vs 4% for predecessor — thegautamkamath · 2026-10-15
- MATS launches paid 6-24 month Residency for independent AI safety researchers — NeelNanda5 · 2026-10-15
- Google ships a Gemini Notebook for post-quantum cryptography migration — philvenables · 2026-10-15
- Vercel's $1M Sandbox challenge drew 1,285 reports and surfaced a Linux kernel zero-day — cramforce · 2026-10-15