Anthropic gives six METR evaluators broad internal access for AI safety reviews

Anthropic announced October 15 that six employees from METR, a nonprofit AI evaluation organization, have begun working with broad internal access. The company said they received equipment, workspace and access to systems comparable to those used by its own risk teams.

The arrangement limits access to customer information, privileged legal material and model architecture details. METR can publish assessments without Anthropic's approval, but Anthropic can require redactions within specified categories. An October 16 clarification said Anthropic decides which information qualifies, while METR retains rights to make certain statements about the access it received.

The move follows Anthropic's September commitment to expand outside scrutiny of its models and safety practices. Anthropic said deeper access helps evaluators assess risks using information that cannot safely be made public.

Anthropic plans to expand its program through three levels of access, with METR occupying the most extensive tier. Other partnerships currently have narrower access and publication rights. METR is expected to publish its findings, though Anthropic gave no release date.