Anthropic Embeds Six METR Evaluators With Internal-Tools Access

Anthropic has onboarded six METR staff with "company hardware, office space, and access to internal tools and systems," a level of access the company describes as "roughly on par with what Anthropic's own risk assessment teams have." The October 15 announcement formalizes a three-tier program for outside evaluators of its frontier models.

Three narrow buckets stay walled off: customer data, attorney-client-privileged information, and model architectures. On publication, "Anthropic cannot veto publication, but it can require METR to redact information" in pre-defined categories, and METR keeps the right to publicly contest any redaction it considers unjustified.

SecureBio sits in the first tier, reviewing chemical and biological weapons sections. Accenture and Faculty are working toward partnerships involving red-teaming and alignment assessments, and Hugging Face is in "early discussions." Anthropic frames the whole structure on financial supervision, writing that "regulators assign resident examiners to work inside the largest banks."

Independence conditions require evaluators to demonstrate expertise, freedom to publish, and that they "should not hold equity in the company they are evaluating." The post arrives in the same week Anthropic is fielding 50-page FTC civil investigative demands over AI safety incidents, one of 342 Anthropic items we have tracked over the past ninety days.