Anthropic’s experiment in outside AI oversight began with a broader pledge. In September, chief executive Dario Amodei proposed embedding independent evaluators inside frontier AI labs, amid mounting concern over agents that can act beyond their intended guardrails. OpenAI later made a similar commitment, but Anthropic is now presenting its arrangement as the first to put that idea into practice.¹

On Thursday, the company said it had brought six researchers from the nonprofit Model Evaluation and Threat Research, or METR, into its operations with company laptops, office space and internal-tool access roughly comparable to that available to Anthropic’s own risk-assessment staff. Anthropic’s case is straightforward: public accountability cannot rest on disclosures that omit the operational details needed to judge a frontier system’s risks. “We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands.”²

METR is meant to supply the independent counterweight. Its embedded team can assess Anthropic’s risk levels, incidents and practices, then publish its own analysis without the company’s approval. The arrangement resembles resident bank examiners, Anthropic says: specialists placed inside institutions whose failures could have public consequences.²

Yet the access is not unlimited. Anthropic cannot veto a METR report, but it can require redactions covering customer information, attorney-client privileged material and model architectures. METR may publicly say whether it considers any redaction justified, provided it does not reveal what was removed. The six researchers are also siloed from the rest of METR, a safeguard Anthropic says protects company, customer and partner information.³

The METR team sits atop a three-tier evaluator system. SecureBio, now in the first tier, receives narrower access and needs approval to publish; middle-tier evaluators gain broader access and publication rights. Anthropic says it hopes more groups eventually reach METR’s employee-like level—voluntarily setting a standard it says every frontier AI company should meet.²