Anthropic gives METR evaluators employee-level access

Anthropic says it has started giving an outside group genuine access to the inside of the company. Six staff from METR, a nonprofit that tests frontier models, now have company hardware, office space, and internal tools roughly comparable to what Anthropic's own risk assessment teams use. Employees are encouraged to talk to them. The only off-limits areas are narrow: customer data, attorney-client privileged material, and model architecture details. Anthropic cannot block what METR publishes, though it can require redactions in a few predefined categories, and METR is free to say publicly whether it thinks a redaction was justified without revealing the content.

This sits inside a three-tier framework. Tier 1 covers narrow investigations that Anthropic approves before publication; Tier 2 lets evaluators publish freely with limited removals; Tier 3 is the METR arrangement, with deep access and independent publication. SecureBio is already reviewing the chemical and biological weapons section of Anthropic's August risk report, and Accenture, Faculty, and Hugging Face are in earlier stages. The comparison Anthropic reaches for is bank supervision, where regulators place examiners inside the largest institutions rather than reading their reports from outside. The company frames this as something every frontier lab should do, not a favor it is granting.

Why it matters

If you follow AI safety claims, the hard problem has always been that the company grading its own homework holds all the evidence. This is a concrete test of whether embedded outsiders with publication rights can close that gap, and whether other labs follow or let Anthropic stand alone.