Anthropic Gives METR Independent Evaluators Employee-Level Access

Anthropic, the company behind the Claude models, has given six employees of METR — an independent nonprofit that assesses safety risks in top AI models — access "like that of a staff employee." That means office space, company equipment, and access to internal tools at roughly the same level as Anthropic's own risk-assessment teams. METR can publish its findings without getting Anthropic's sign-off. Anthropic can ask for redactions only in three narrow categories: customer data, materials protected by attorney-client privilege, and details of model architecture. This follows a commitment Anthropic made last month to strengthen independent third-party oversight, made under the White House Accord on Superintelligence — an agreement among AI companies to work with independent auditors.

Anthropic cannot veto findings it does not like. It also limited its own editing power to those three narrow categories. Both of these are unusual for the industry. Normally the public has no choice but to take a lab's safety claims on faith. Here an outside group can publish things that embarrass Anthropic, and it does not need permission to do so. It is not yet clear whether other AI labs will follow with the same level of access, or whether this will remain a one-off step by Anthropic.