Story
Oct 16, 2026
Anthropic Invites METR Inside—but Keeps a Hand on the Redaction Lever
Anthropic has given six nonprofit METR researchers unusually deep access to its systems and the right to publish risk findings. The company says the arrangement strengthens accountability, while retaining narrowly defined powers to force redactions.
Last month, amid mounting concern over increasingly capable AI systems and calls for tougher outside checks, Anthropic committed to stronger third-party evaluation and signed the White House Accord on Super Intelligence. The company argued that public accountability requires independent scrutiny, but said it could not disclose every safety detail because some material could aid malicious actors or expose commercially sensitive information.1
On Thursday, Anthropic turned that pledge into a concrete experiment: six employees from the nonprofit Model Evaluation and Threat Research, or METR, were onboarded with company hardware, office space and internal-tool access “roughly on par” with Anthropic’s own risk-assessment teams.1 The company says staff should share information with the embedded researchers much as they would with an employee working on alignment risk.
Anthropic’s case is that this is deeper than a conventional audit. Drawing a comparison with resident bank examiners, it says the goal is to narrow the information gap between a frontier AI developer and those judging its risks. METR will assess the company’s risk levels, incidents and practices, then publish its own analysis.1
Yet the independence on offer has clear boundaries. Anthropic cannot veto publication, but it can require redactions covering customer data, attorney-client privileged information and model architectures. METR may publicly say whether it regards a redaction as justified, provided it does not reveal what was removed.2 That makes the arrangement unusually transparent by industry standards, but not wholly free of company control.
The METR team is currently alone in Anthropic’s top tier of evaluators. SecureBio, reviewing the chemical and biological weapons section of Anthropic’s August risk report, has narrower first-tier access and needs approval to publish; Anthropic is also pursuing partnerships with Accenture, Faculty and potentially Hugging Face.1
Anthropic calls the program voluntary and says every frontier AI company should follow. For now, it is a test of whether an evaluator can look sufficiently far under the hood while the company being examined still holds the redaction key.1