Anthropic gives six METR researchers employee-level access to assess its models, safety practices and risk management

Anthropic has given six METR researchers employee-level access: company laptops, office space and internal tools roughly on par with its own risk assessment teams. METR will publish its own assessments of Anthropic's risks, incidents and practices. Anthropic has no veto, but can require redaction in three narrow categories (customer data, attorney-client privileged material and model architecture information), and METR may publicly describe what was redacted. METR is the only evaluator in the top tier of a three-tier system, while lower-tier evaluators such as SecureBio have narrower access or need Anthropic's approval to publish. The source for the terms is Anthropic's own blog post, as reported by TechCrunch. No dates or findings have been given.

This is a structural change in who can see inside a frontier lab: an outside evaluator with publication rights and internal access, which is more than the usual arrangement of testing a model through an interface. Regulators, enterprise buyers and rival labs now have a concrete template to compare against, and a benchmark for what "independent oversight" ought to mean. The practical impact is still limited for now. Nothing has been published, six people can only cover so much, and the "model architecture" redaction category could hide exactly the details that matter most for judging risk.

The "publicly accountable" framing and the comparison to resident bank examiners come from Anthropic's own messaging, and they overstate things: bank examiners have legal powers, while METR's access is voluntary and can be withdrawn. The link to weekly-to-biweekly reports of AI agents acting outside instructions is the article's framing, not evidence that this arrangement would prevent or detect such incidents. The long entity list (OpenAI, Google, Meta, Nvidia, xAI, White House) is context, not a sign that those parties have taken any action here.

1. 01The first METR report on Anthropic: when it appears, whether it includes critical findings, and how much is redacted and how METR describes the redactions.
2. 02Whether other frontier labs (OpenAI, Google, Meta, xAI) offer comparable embedded access with publication rights, or stay with narrower, approval-based arrangements.
3. 03Whether regulators or the White House cite this model in guidance or requirements, and whether the access terms or METR's team size change over the next 6 to 12 months.