AI news · October 16, 2026
Anthropic gives METR employee-level access for independent evaluation
Anthropic will give six METR employees access to internal tools and company hardware, while preserving publication rights, making independent agent evaluation more inspectable.
- METR employees
- 6
- evaluation groups named
- 3
Anthropic announced on October 15 that six employees from METR will work in its offices with company hardware, internal tools and access close to the risk team. METR can publish its findings. Anthropic says it can ask for narrow redactions only for customer data, attorney-client material or model architecture. The update also names SecureBio and Accenture as independent evaluators, with different access arrangements.
The important change is access to the records and settings around a model, not another public score. Outside researchers can see more of the path from a model prompt to an agent action, while Anthropic keeps a limited review for sensitive material. This is still not the same as a fully open audit: the scope, results and redactions will decide how useful it becomes. Builders should watch the reports and copy the evidence trail, not just the headline partnership.
What you can do with it
Read METR's future reports and copy the method for one agent task: keep the prompt, tools, permissions, model version, traces and failed attempts. Ask an outside reviewer to repeat the task without access to your private data, then compare the findings with your internal score.
Our take
This is a meaningful governance move because the evaluator gets close to the operating system of the model, not only a test endpoint. Its value depends on published work and clear scope. Until those arrive, treat access as a promise to inspect, not evidence of safety.
Links AnnouncementCoverage
Source: Anthropic — Made With Models writes the brief; the reporting is theirs.