Anthropic gives outside AI watchdog “employee-like” access

Anthropic says it has onboarded six people from METR, a nonprofit that evaluates frontier AI models, giving them company hardware, office space, and access to internal tools roughly on par with its own risk assessment teams. Staff are encouraged to share information with them as they would with a colleague working on alignment risk.

The key difference from past arrangements is publication. METR will publish its analysis of Anthropic's risk levels, incidents and practices, and Anthropic says it cannot veto that. It can require redactions in predefined categories like customer data, privileged legal information and model architecture details.

This is the first concrete step on a pledge that drew rare agreement across rival lab leaders in September, and it comes as Anthropic heads toward an IPO that Ars Technica, citing the Financial Times, reports is expected at a valuation above $2 trillion. If outside eyes inside frontier labs become normal, this is what the template looks like.

We think this is a real step and deserves credit as one. For years, outside AI safety testing has meant labs handing evaluators a model and a deadline. This is different. Embedded people with internal access can see how models were trained and how safeguards actually work, not just what a finished model does on a test. Anthropic's own comparison is to bank examiners who sit inside the largest banks. That is the right instinct for an industry whose failures could land on the public.

The independence details are better than we expected. Anthropic says it is not paying METR, and Ars reports METR takes no donations from AI companies or their staff, relying instead on individual and philanthropic donors and government bodies such as the UK's AI Security Institute. Anthropic says the embedded team operates under enforceable agreements and is walled off from other METR researchers. Those are the bones of a credible watchdog.

But look closely at the redaction rules. Anthropic's post, updated a day after launch to clarify METR's terms, says Anthropic ultimately decides what falls into the protected categories, while METR keeps the right to describe the access and information it received. So there is no veto, but there is a redaction pen, and the company holds it. That matters. A report with the important parts blacked out, plus a polite note about the blacking out, is not the same as oversight.

Critics quoted by Ars raise the deeper problem: conflicts of interest, with accounting as the cautionary tale, and the view that this should be government regulation, not a company's choice. We share that worry. The bank analogy cuts against Anthropic here. Resident bank examiners work for regulators with legal power. METR works under an agreement Anthropic drafted and can, in principle, end.

Where this leads if it scales: if every frontier lab embeds independent evaluators with publication rights, the public gets far better evidence about risk than it has today, and that would be good for useful AI too, since trust is a precondition for adoption. If it stays voluntary and uneven, it becomes a credential labs can point to while setting the terms themselves. OpenAI, per Ars, has not yet laid out its own plan. I'm watching METR's first published report, how much of it is redacted, and whether anyone else follows.

Anthropic describes three tiers of evaluator access. First-tier evaluators run narrow investigations and need approval to publish. Second-tier evaluators get more information and can publish without approval, subject to redactions they can publicly contest. The third tier, so far only METR, adds employee-like access. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, the company is building a partnership with Accenture and Faculty, and it is in early talks with others including Hugging Face. Anthropic also signed the White House Accord on Super Intelligence, which included a pledge to work with independent auditors or evaluators.