Anthropic's embedded METR evaluator program

Anthropic says six METR employees have been embedded with company hardware, office space and access to internal risk-assessment tools and systems. METR can publish analysis of risks, incidents and practices without an Anthropic publication veto, subject to narrowly defined redaction categories. It can characterize those redactions publicly without disclosing the underlying protected information.

The arrangement is the highest of three access tiers described by Anthropic. It applies to a limited team under enforceable agreements, with other METR staff siloed from that work. These are announced oversight terms, not completed independent findings or proof of effective prevention. Other evaluator partnerships have different access and publication rights.