Anthropic Gives METR Employee Level Access for Safety Reviews

Anthropic has implemented its plan to embed outside AI safety evaluators, giving six employees of METR access to company hardware, offices and internal systems, Reuters reports. Their access is roughly equivalent to that of Anthropic's risk assessment teams.

METR cannot access customer data, legally privileged information or model architecture. Anthropic cannot veto its findings about risks, incidents and company practices, but can require narrowly defined redactions. METR can publicly challenge those redactions without disclosing the withheld information.

Anthropic is also working with SecureBio and Accenture's Faculty business under arrangements with more limited access and publication rights. SecureBio is reviewing the chemical and biological weapons section of Anthropic's August risk report, while Faculty is working on model evaluations, adversarial testing and assessments of model behavior.