Anthropic Gives METR Employee-Level Access to Audit Its AI Safety

Anthropic said Thursday it has onboarded six researchers from METR, a nonprofit that evaluates frontier models, with employee-like access to its systems. The team gets company laptops, office space, and internal tools roughly matching what Anthropic’s own risk-assessment staff use. Outside narrow categories such as customer data, attorney-client material, and model architecture details, staff are told to share with them as they would with a colleague on alignment risk.

METR will publish its own read on Anthropic’s risk levels, incidents, and practices. Anthropic cannot block those reports, though it can force redactions in those predefined buckets; METR can say publicly whether it thinks a redaction was justified. The company put METR alone at the top of a three-tier evaluator system that also includes SecureBio and Accenture.

The setup fulfills a pledge CEO Dario Amodei floated last month and tracks a White House accord on independent review that Anthropic, OpenAI, Google, Meta, Nvidia, and xAI signed.

— via Anthropic