Anthropic gives six METR evaluators employee-level access to internal systems

What is changing?

Anthropic said on October 15 that it had onboarded six people from the independent AI evaluation nonprofit METR with company hardware, office space, and access to internal tools broadly comparable to its own risk assessment teams. A limited set of categories remains out of scope, including customer data, attorney-client privileged information, and model architecture details.

METR will be able to publish analysis of Anthropic’s risk levels, incidents, and practices. Anthropic cannot veto publication, although it may require redactions for predefined sensitive information; METR may publicly discuss whether those redactions were justified without revealing the withheld material. Anthropic is building a three-tier evaluation model ranging from narrow, company-approved publication to employee-level access.

Reuters notes that independence will be judged not only by depth of access but also by who selects the evaluators and how intellectual-property restrictions affect scrutiny of model risks. For production teams, embedded evaluation could provide evidence beyond model cards, but credibility will depend on transparent documentation of access limits, redactions, and publication rights.