Anthropic Gives METR Employee-Like Access to Internal AI Systems

What changed

Anthropic has embedded six METR employees in its offices with access to company hardware, internal tools and systems used by its own risk-assessment teams. METR can publish its findings without Anthropic’s veto, though the company may require redactions for customer data, attorney-client privilege and model-architecture details.

Why it matters

This changes an outside audit from a scheduled inspection into something closer to a resident risk team. METR’s evaluators can see more of the machinery while it is being used, not just the polished results Anthropic chooses to present.

That access is meaningful only if independence survives proximity. Anthropic is not paying METR, and the nonprofit says it does not accept donations from AI companies or their employees. Its publication rights also give the arrangement a sharper edge: the evaluator can make uncomfortable findings public.

There is still no timetable for evaluations or publication, and no measurable finding has been announced. So the important result today is structural, not scientific. Anthropic has created a channel through which external scrutiny can reach internal systems, while keeping narrow limits around sensitive information.

The company is also continuing work with Accenture and SecureBio on safety testing. The next useful test will be whether METR’s access produces findings that materially challenge Anthropic’s own risk assessments, rather than simply validating them.