Anthropic gives METR inside access and no veto

Six people from METR now work with Anthropic company hardware, office space, and internal access roughly on par with Anthropic's own risk assessment teams. That is the core of Anthropic's update on independent evaluators, published October 15, 2026. The detail that matters most is the publication clause. METR can publish its analysis of Anthropic's risk levels, incidents, or practices, and Anthropic cannot veto it. Anthropic can require redactions in narrow predefined categories (customer data, privileged legal material, model architecture). METR is free to say publicly whether it thinks those redactions were justified.

That last clause gives the arrangement value beyond public relations. A review that the company scopes and approves, which is how Anthropic describes its own first tier of evaluator access, tells you what the company was willing to show. In the third tier, the evaluator decides what goes out, and the company has to defend its redactions in public. Anthropic compares this to resident bank examiners, and the comparison holds as far as it goes. Bank examiners, though, sit on a statute. Here the access is voluntary, it rests on Anthropic's agreements rather than law, and six people is a small team set against a frontier lab's whole model pipeline.

From Paris, this reads like a private rehearsal for what the EU AI Act already asks of providers of general-purpose models with systemic risk: model evaluation, adversarial testing, incident reporting. If embedded evaluators can work for a year without leaking anything sensitive, regulators on both sides of the Atlantic get a working template. The question I'd want answered is what happens to METR's access the first time a finding is genuinely embarrassing.