Anthropic Grants Employee-Level Access to Independent AI Auditors, Drops Veto Power

Anthropic is granting six researchers from the safety nonprofit METR employee-level access, including company laptops and internal tools, to independently evaluate its frontier models. The company has surrendered its right to block METR from publishing its risk and safety assessments. Anthropic retains narrow redaction rights covering customer data, model architecture, and privileged legal information, though METR is explicitly authorized to publicly challenge the justification of any redactions. The arrangement places METR at the top of Anthropic’s new three-tier evaluator system, distinct from lower-tier auditors like SecureBio that still require publication approval. "We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands." — Anthropic