Anthropic gives METR deep access for safety audits

Artificial intelligence developer Anthropic has embedded six staffers from the non-profit research group METR within its safety evaluation setup, providing them with corporate hardware, physical office space, and access to internal tooling. The company stated that this access places the outside auditing team roughly on par with its own internal risk groups.

While the arrangement provides broad exposure to Anthropic's safety protocols, certain boundaries remain in place. METR will not have access to customer data, material covered by attorney-client privilege, or core model architecture details. However, METR retains the right to publish its evaluation results independently. Anthropic can only request redactions within predefined categories, and METR remains free to publicly disclose whether it believes those redactions were fair.

The initiative represents a significantly deeper level of access than Anthropic has granted to other external evaluators. Firms like SecureBio, Accenture, and Faculty have previously received more restricted access alongside tighter limits on what they could publish. Anthropic noted that it has also entered early talks with Hugging Face, coming on the heels of signing the White House Accord on Super Intelligence, which commits signatories to external safety reviews.

For AI governance researchers and industry practitioners, the deal provides a functional template for third-party safety audits that balances commercial confidentiality with meaningful oversight. By granting external teams direct tool access while allowing public commentary on redaction disputes, the framework establishes a higher standard for independent algorithmic auditing.