METR proposes independent AI risk reviews, with labs controlling access and redactions

The proposal would examine agent behavior, monitoring and AI-assisted research. It seeks public evidence, not a safety certification, and leaves developers substantial control over what assessors can see.