Anthropic has given six employees of METR, a nonprofit that evaluates frontier AI models, access to internal tools and systems similar to those used by its own risk-assessment teams. In its announcement, Anthropic said the researchers had also received company hardware and office space.
The arrangement brings outside evaluators closer to Anthropic’s internal safety work. Employees are encouraged to share information with the METR team much as they would with colleagues assessing AI alignment risks.
METR can publish assessments of Anthropic’s risk levels, incidents and practices without seeking the company’s approval. Anthropic can still require redactions involving customer data, attorney-client privileged material and model architectures. The company ultimately decides which information falls into those protected categories.
Publication rights have limits
METR retains the right to make certain public statements about the information and access it received. However, the six embedded employees work under enforceable agreements, and other METR employees do not have access to the embedded team’s work.
Anthropic describes three levels of access for external evaluators. At the first level, the company decides what information to provide and must approve publication. At the second, evaluators receive wider access and can publish without approval, subject to certain redactions. METR occupies the third level, combining employee-like access with publication rights similar to those at the second level.
Anthropic has worked with METR on reviews of its Risk Reports for about a year. It is also working with SecureBio on chemical and biological weapons risks and developing a broader evaluation partnership with Accenture’s specialist AI business, Faculty. Those partnerships have not reached METR’s level of access.
The expanded access follows Anthropic’s September commitment to stronger independent evaluations. Under the voluntary White House Accord on Super Intelligence, signatory companies pledged to work with independent external assessors to examine whether their internal AI safeguards operate as intended.
What independent evaluation can establish
METR has separately warned about the limits of voluntary risk assessments. In a proposal describing potential arrangements with AI developers, the nonprofit said companies could refuse requests for internal information, leaving outside assessors with less access than senior safety employees. It also said the proposed exercises would not constitute audits or compliance assessments.
The nonprofit’s separate statement on independence said it does not accept payments or donations from frontier AI developers. Even without direct funding ties, METR acknowledged that its work depends on companies voluntarily providing access and allowing information to be published. It aims to disclose limitations that could affect how readers interpret its findings.