An update on independent evaluators

Anthropic is widening the circle around its model safety work. After promising stronger independent review last month and signing the White House Accord on Super Intelligence, the company says it has now given a team from METR, the nonprofit AI evaluator, access that looks a lot like being inside the building.

That access is not symbolic. Anthropic says it has onboarded six METR employees with company hardware, office space, and internal tools and systems that are roughly on par with what its own risk assessment teams can use. The arrangement lets Anthropic share information with them much as it would with its own staff, except for a few defined categories, including customer data, attorney-client privileged material, and model architecture details.

The logic is simple, and a little uncomfortable for AI companies: the people judging the risk need to see enough to judge the risk. Anthropic argues that the public cannot be shown everything that matters, because some of it could help malicious actors or expose sensitive training and compute information. So it is letting outside evaluators into the process instead, with the promise that they can publish evidence and reasoned conclusions about what they find.

METR is the furthest along. Anthropic says the group will publish its analysis of risk levels, incidents, or practices, and Anthropic cannot block publication, though it can demand redactions in narrow categories. METR can also publicly say whether it thinks those redactions were justified. This is meant to feel less like a consultant relationship and more like supervision, which Anthropic explicitly compares to bank examiners working inside large banks.

Other partnerships are still earlier and narrower. Anthropic says it is working with SecureBio and Accenture, and is in early discussions with several other organizations, including Hugging Face. It is also laying out three tiers of evaluator access: first, tightly scoped investigations with company approval for publication; second, broader access and publication without approval; and third, the METR-style setup with employee-like access. SecureBio is currently reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report after a pilot review of a previous report, while the work with Accenture and Faculty is aimed at model evaluation, red-teaming, alignment assessments, safeguard testing, and capability building.