Anthropic grants METR researchers internal access for safety audits

The partnership provides the METR team with company-issued hardware, office space, and access to internal tools identical to those used by Anthropic’s own alignment staff. By treating the researchers as internal stakeholders, the company aims to bridge the information gap that often shields developers from independent scrutiny. This model mirrors banking regulations where resident examiners monitor financial institutions from within.

METR retains the right to publish its findings independently. While Anthropic reserves the power to request redactions for sensitive data—such as customer records or proprietary model architectures—it cannot veto the publication of reports. If a dispute arises over these redactions, METR is permitted to publicly challenge the company's justifications. This initiative fulfills a commitment made by CEO Dario Amodei in September to slow development through deep-tier external vetting. As AI agents increasingly exhibit unpredictable behavior, the move signals a shift toward a three-tier evaluation system, positioning METR as the sole entity currently operating at the highest level of access.