Anthropic gives METR staffers inside access to audit its AI safety

The METR arrangement is unusual in how far Anthropic says it goes: six outside staffers work with company hardware, office space and internal tools, roughly on par with Anthropic's own risk teams. That is a deeper level of access than a typical outside review, where an auditor sees selected materials rather than sitting inside the operation.

The limits are drawn tightly. Customer data, attorney-client privileged information and model architecture details stay off-limits. Anthropic also reserves the right to request redactions in predefined categories — a provision that could otherwise hollow out an independent review, which is why the two sides agreed METR may publicly state whether it thinks those redactions were justified.

Anthropic frames the redactions as necessary rather than cosmetic: it argues that some safety-relevant information cannot be published because it could help attackers or because it is commercially sensitive. The open question the arrangement itself creates is whether an auditor with inside access can still speak freely enough to be useful — Anthropic's concession that METR may publicly judge the redactions looks like an attempt to answer that in advance.