Anthropic gives METR employee-like access, gives up veto

Six METR employees get tools roughly on par with Anthropic’s own risk teams, and the company can’t veto what they publish, Anthropic said.

Anthropic said Thursday it has started giving “employee-like access” to a team from METR, a nonprofit that evaluates frontier AI models, according to a company post.

METR will publish its own analysis. Anthropic said it can’t veto publication.

The rule in English

Anthropic said it onboarded six METR employees. They get company hardware, office space and internal tools and systems “roughly on par” with what Anthropic’s own risk assessment teams have.

Anthropic employees are encouraged to share information with them as they would with a fellow employee, the company said. Exceptions are a small number of defined categories. Examples include customer data, attorney-client privileged information and model architectures.

Anthropic said it can require redactions only in narrow, predefined categories. METR may say publicly whether it thinks those redactions were justified, as long as it doesn’t reveal the redacted material.

Who wrote it

Anthropic set the terms. It said it is embedding evaluators voluntarily. Last month the company said it committed to stronger third-party assessment. It also signed the White House Accord on Super Intelligence, in which companies pledged to partner with outside auditors or evaluators.

Who it binds

The METR team works under “enforceable agreements,” according to Anthropic. Other METR employees are siloed from that team’s work.

Who it exempts

Other evaluators get less, for now. Anthropic said it is building three tiers of access. METR sits at the third tier. SecureBio is in the first tier, where Anthropic decides access and approves publication. Anthropic said it is also working with Accenture and Faculty and is in early talks with others, including Hugging Face.

Anthropic said it believes every frontier AI company should embed evaluators.

What’s next

METR will publish the results of its risk assessments, Anthropic said. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report. Anthropic said it hopes to move more evaluators to the third tier and said it is committed to sharing updates.