Anthropic Gives METR Employee-Level Access

Anthropic is giving researchers at METR employee-level access to study its models, risks, incidents, and internal practices. The arrangement is unusual because METR will publish its own assessments, and Anthropic has reportedly given up the right to block those publications. That combination is meant to strengthen independent scrutiny of frontier AI development. External audits have become a central demand among policymakers, safety researchers, and critics who worry that companies cannot credibly evaluate their own systems. Employee-level access could let METR examine not only model outputs but also the context around development decisions, safety testing, and incident response. The ability to publish without approval is perhaps the most significant part of the deal. It creates a test of whether a major AI lab can open itself to outside experts while still protecting trade secrets, personal data, and security-sensitive information. If METR produces useful, fair, and technically rigorous reports, the model could encourage similar agreements across the industry. If the process becomes too limited or too contentious, it may show the limits of voluntary transparency. The arrangement also raises questions about how independent METR can be when it relies on access granted by the company being studied. Still, the move is a concrete step beyond public model cards and blog posts. It may set expectations for how AI firms allow outside experts to examine their internal work, especially as governments consider mandatory evaluation regimes. The results will be watched closely by both safety advocates and AI companies.