Anthropic extended an arrangement on 15 October that it began earlier this year, when it first paid outside evaluators to work inside its own building. Six staff from METR, a nonprofit that assesses frontier models, now have company hardware, office space and access to internal tools roughly comparable to Anthropic’s own risk assessment teams. Employees are told to share information with them, except for customer data, privileged legal material and model architecture.
The publication terms are the load-bearing part. METR can publish its risk assessments, and Anthropic cannot veto publication, though it can require redactions in narrow predefined categories. METR is then free to say publicly whether it thinks those redactions were justified, without revealing what was cut. An evaluator who can be silenced is a reviewer in name only, so that clause is what separates this from a courtesy review.
Anthropic also published a three-tier framework: Tier 1 covers narrow investigations where it approves publication, Tier 2 allows publication without approval, and Tier 3 is the METR arrangement. SecureBio is reviewing one chapter of the August risk report. Anthropic says every frontier company should do this, which is a claim about others rather than a commitment anyone else has made. Read the full update on Anthropic’s site.