What changed
Anthropic says six employees from the nonprofit evaluator METR have begun receiving employee-like access to internal tools, systems, and working space. The arrangement gives them access beyond an ordinary outside evaluation, while excluding defined sensitive categories, including customer information, legally privileged material, and model-architecture details.
Who can publish
According to Anthropic, METR can publish its analysis without an Anthropic veto. Anthropic can require redactions of predefined sensitive information; METR can publicly describe those redactions and its objections without revealing the protected material. The announcement describes access and publication rules, rather than a completed independent assessment. We have checked the announcement, not audited the arrangement.
Our take
Editorial analysis: the useful questions are what evaluators can inspect, whether they can report uncomfortable results, and what readers cannot see. Broader access could make a review more informative, but that remains to be demonstrated in published findings. Treat an evaluation partnership as a way to seek evidence, rather than a certification that a model is safe.