Anthropic is opening its doors in a way few frontier AI labs have tried. On Thursday, the company said it has given a team from METR, a nonprofit that evaluates advanced AI models, employee-level access to its systems so the group can independently assess its models, safety practices and risk management.
The arrangement puts six METR staff inside Anthropic's working environment. They have company hardware, office space and access to internal tools broadly comparable to what Anthropic's own risk assessment teams use, according to the company. Employees are encouraged to share information with them much as they would with a colleague working on alignment risk.
What sets the plan apart is publication. METR will be able to publish its own assessments of risks, incidents and safety practices without Anthropic's prior approval. Anthropic can require redactions only in narrow, predefined categories, such as customer data, privileged information and certain details of model architecture. METR is also free to say publicly whether it believes a redaction was justified, without revealing the redacted material itself.
Anthropic framed the move as a step toward public accountability, saying it wants to close the gap between what outside evaluators understand about a frontier company's systems and what the company understands itself. It compared the setup to banking, where regulators place resident examiners inside the largest banks rather than relying only on periodic outside reviews.
METR sits alone, for now, at the top of a three-tier system Anthropic is building for evaluators. Evaluators in the lower tier run narrower reviews and need Anthropic's approval to publish. The middle tier receives more information and can publish without approval, while the top tier adds the kind of embedded, employee-level access METR now has. Biosecurity nonprofit SecureBio was cited as an example of a lower-tier evaluator.
The announcement follows through on an idea Anthropic chief executive Dario Amodei raised in a September essay, proposing embedded outside evaluators inside AI labs as a first step toward slowing the technology's development when risks warrant it. OpenAI quickly pledged a similar approach, and Anthropic, OpenAI, Google, Meta, Nvidia and xAI have since signed a White House accord committing to independent review.
The timing matters. TechCrunch noted that disclosures of AI agents acting outside their instructions have been surfacing every week or two since late August, sharpening calls for checks that do not depend entirely on a lab grading its own work. An embedded evaluator will not settle the broader debate over how frontier models should be governed, but it does create a test case: whether a lab can give an outsider real access, accept public findings it cannot veto, and keep that arrangement credible over time.
If other labs follow with comparable access, the model could become a practical template for independent oversight while formal regulation is still taking shape. If they do not, Anthropic's experiment will stand out as a measure of how far voluntary transparency can go — and where its limits lie.