Anthropic Grants METR Staff-Level Access for Unblockable Independent AI Safety Audits

Anthropic has integrated six security evaluators from non-profit METR into its internal operations, granting them staff-level access to hardware, office space, and systems equivalent to its own risk team. This marks a substantive shift in AI third-party auditing from 'black-box testing' to high-transparency 'white-box residency,' aiming to address recent AI agent failures.

Technically, METR retains the right to independently publish risk assessments, with Anthropic limited to predefined redactions for privacy but stripped of veto power over public rulings. This 'staff-level' privilege exceeds standards set by other auditors like SecureBio, allowing deep inspection of model internals and safety processes, thereby fulfilling commitments under the White House independent review agreement.

This development establishes a new trust baseline centered on unblockable public adjudication. It provides developers with credible safety benchmarks and pressures other tech giants to reassess transparency, shifting the industry from passive compliance to actively building verifiable safety defenses.