Anthropic gives METR employee-like access to audit frontier AI safety

Anthropic says it has started giving employee-like access to a team from METR, a nonprofit that evaluates frontier AI models — a concrete step after last month’s pledge on stronger third-party assessment and the White House Accord on Super Intelligence. The details are in Anthropic’s own post.

What “employee-like access” means

Anthropic has onboarded six METR employees with company hardware, office space and access to internal tools and systems that it says are roughly on par with what its own risk-assessment teams use. Outside narrow categories — customer data, attorney-client privileged material, or information about model architectures — Anthropic staff are encouraged to share with them as they would with a colleague on alignment risk.

METR will publish analyses of Anthropic’s risk levels, incidents or practices. Anthropic cannot veto publication, but it can require redactions that fall into those predefined categories. METR may say publicly whether it thinks a redaction was justified, without revealing the redacted content.

The engagement covers a small embedded team under enforceable agreements; other METR staff are siloed from that work.

Why Anthropic wants outside eyes

Anthropic argues that as models grow more capable, risk assessments get harder — and companies cannot publish every sensitive detail. Embedding evaluators who can see training and safeguard context, then publish reasoned conclusions, is meant to narrow the gap between what the company knows and what outsiders can verify. Anthropic compares the idea (imperfectly) to resident bank examiners.

Three tiers and other partners

Anthropic describes three access tiers, from narrowly scoped reviews that need approval to publish, through publication without approval with more information, up to the METR-style embed. It is also working with SecureBio (on chemical and biological risk chapters of its Risk Report) and toward broader work with Accenture and Faculty, and says it is in early talks with others including Hugging Face.

For anyone watching AI governance, the story is less a new model and more a test of whether deep, publishable third-party access can stick inside a competitive lab.