Anthropic and OpenAI commit to embedding third-party safety evaluators
Following Amodei’s proposal, Anthropic and OpenAI say they will embed independent evaluators (e.g. METR, Redwood) with access to report incidents and assess alignment. Outside researchers welcome the idea but warn NDAs, short access windows, and missing details could leave them as contractors rather than true watchdogs; Meta, SpaceXAI, and DeepMind have not committed.
