Signal Frontier Risk Report (Feb–Mar 2026) — METR

Summary

METR published its Frontier Risk Report on 19 May 2026, covering a pilot conducted from late February through mid-March 2026 with Anthropic, Google, Meta and OpenAI. The pilot was an entity-based (not model-specific) assessment of loss-of-control risks from internal use of AI within frontier developers, designed to be repeated periodically rather than tied to individual model launches. Each participant gave METR access to its most capable internal models, including raw chains of thought, plus non-public information on capabilities, internal AI use and monitoring, and pace-of-progress trends. METR wrote private reports for each participant, who then approved which non-public details could be disclosed before the public report was finalized; the resulting report spans more than 300 pages and is described as the first independent, cross-industry assessment of loss-of-control risks from advanced AI agents. Anthropic separately contributed red-teaming findings included in the report's appendix, and METR called for periodic third-party assessment of this kind to be adopted industry-wide.