Anthropic puts six outside evaluators in its offices and gives up the right to block their reports

A company-issued computer, a desk on its premises, and access to internal tools comparable to what its own risk evaluation teams have. That is what Anthropic, the developer of the Claude AI models, has given six employees of METR, a nonprofit that evaluates the most advanced AI models. The company announced it on October 15 in a blog post.

The novelty lies mainly in the publication rights. METR will publish its own analyses of Anthropic's risks, incidents and practices, and the company gives up the ability to block them. It can only demand redactions in categories defined in advance, such as customer data, information covered by attorney-client privilege, or model architecture. METR will be able to say publicly whether it considers those redactions justified, without revealing what they contain.

Why now?

The announcement comes after several weeks of incidents. According to TechCrunch, AI agents straying from their instructions have been reported every one to two weeks since late August. Anthropic itself cut off Internet access for its evaluations after unexpected actions by Claude, and the European Commission asked both companies for explanations. Anthropic also notes that it signed an agreement with the White House last month, committing to work with independent outside auditors or evaluators.

The framework Anthropic describes has three tiers. In the first, the evaluator carries out a targeted investigation and needs the company's approval to publish. That is the case for SecureBio, which is reviewing the chemical and biological weapons chapter of Anthropic's latest risk report. In the second, the evaluator publishes freely, subject to the agreed redactions. The third adds access “like an employee.” Only METR has obtained it so far. Faculty, Accenture's AI subsidiary, is working on a broader partnership, and talks are under way with Hugging Face.

Truly independent evaluators?

Several questions remain open. According to Reuters, experts worry about the independence of these evaluations, since the companies themselves choose their evaluators. The news agency also reported earlier this week that Anthropic and OpenAI were seeking to keep substantial control over what outside evaluators could publish, for fear of trade secret leaks. Reuters also points to a fundamental difficulty: the information a company considers strategic is sometimes exactly what matters for assessing a risk.

Anthropic's post does not say who pays for the METR team, and it gives no date for its first report. The company also acknowledges that it cannot share everything publicly. According to Reuters, METR says its report will describe the scope and limits of its access. Finally, TechCrunch reports that OpenAI has committed to the same approach of outside evaluators based in its offices, though the terms are not known.

What to watch

For an IT leader deploying these models, the appeal is simple: one day having a risk assessment that does not come from the vendor itself. Everything will hinge on METR's first report, on how much text was redacted, and on what METR says about it.