Anthropic opens the door to METR: outside evaluators get employee-level access, and the key is who gets to publish

By BankInfoSecurity · October 16, 2026.

Anthropic will bring independent evaluators into its safety process. According to the outlet, staff from METR (an organization that evaluates risks from frontier AI models) will have "employee-like access" to models before deployment, as Anthropic itself explained in a blog post.

The arrangement, as described, is specific: six METR employees with access to hardware, offices and internal tools "roughly on par" with Anthropic's risk-evaluation teams. Anthropic staff are encouraged to share information with them, except for "a small number of clearly defined categories," such as customer data. The METR team inside Anthropic will be small and walled off from the rest of METR.

The key clause concerns publication. METR will publish its own analysis of Anthropic's risk levels and practices. According to the article, Anthropic cannot veto it and can only request that certain information be redacted.

METR details what it needs to assess alignment (whether a model does what its developers intend). It describes two phases: surveys of observed incidents and tests in simulated hypothetical scenarios, to check whether the training process might be masking signs of misalignment. To do so, it requests access to training checkpoints, transcripts and datasets, code and infrastructure, staff interviews and API access.

The cost helps gauge the scale of the challenge. METR says that while investigating the incident between Hugging Face and OpenAI, it spent about $400,000 in API credits over six days, and estimates that a rigorous two-month investigation of that scope would cost more than $10 million. In that case, according to the article, neither METR nor its partner Redwood Research had direct access to OpenAI's infrastructure.

Anthropic also proposes a three-tier system. In the first tier, evaluators conduct limited-scope investigations and Anthropic decides what they can see and publish. The second offers more information, but Anthropic retains the right to withhold categories, and evaluators can publish their view on whether the redactions were justified. The third, where METR now sits, provides employee-level access. Anthropic says other partnerships are at very early stages, with more limited information and publication rights.

Our take: the premise is that trust in AI safety cannot rest on labs' self-reporting, and this move is a step in the right direction. An evaluation without access to training checkpoints or infrastructure is, in practice, an evaluation of the facade. Veto-free publication rights are what turn access into accountability.

Still, a caveat is in order. This is six people, in a single tier-three relationship, and Anthropic itself acknowledges that other evaluators will have less access and fewer rights. Moreover, the line on what gets redacted remains Anthropic's call. The $10 million figure points to another problem: thorough auditing is expensive, and only a handful of organizations can afford it.

More broadly, commitments of this kind come amid political pressure: according to the article, Anthropic and OpenAI signed a commitment at the White House that includes potential partnerships with independent researchers. The tiered model is likely to become a de facto standard, with the risk that the top tier ends up reserved for the evaluators closest to the lab.

This is the kind of institutional infrastructure we need in the short term, as more capable models arrive and the risks are real. In the long term, the promise of AI that helps eradicate disease and generate abundance is only credible if we can verify, through outside eyes, that these systems do what they claim. The test will be METR's first report and how much of it is redacted.