For months the debate about who checks the AI labs has been stuck on an obvious weakness: the outside evaluators meant to vet frontier models have little real access to them. On Thursday, Anthropic tried to answer that criticism in the most literal way possible. It announced that six researchers from METR, the nonprofit best known for investigating a breach by OpenAI's agents, now have company hardware, office space and internal tools at Anthropic, with a level of access the company says is roughly on par with its own risk-assessment staff. The watchdogs, in other words, have been given desks.
The arrangement is more carefully built than a publicity gesture. Anthropic says its employees are encouraged to share information with the METR team as they would with a colleague, with a few carved-out exceptions: customer data, attorney-client material, and the architecture of its models. METR can publish its findings, and Anthropic cannot veto that publication. The company can require redactions in predefined categories, but METR is free to tell the public that a redaction happened and whether it thinks the censorship was justified. To keep the access from leaking outward, the six embedded researchers are walled off from the rest of METR.
Anthropic frames this as the top rung of a three-tier ladder. At the first tier, an evaluator runs narrow investigations and publishes only with the company's sign-off. At the second, it can publish without approval and sees more. The third tier, the one METR now occupies, combines insider access with those stronger publication rights. The company says it is also working with SecureBio, which is reviewing the chemical and biological weapons section of its latest risk report, and with Accenture and Faculty on red-teaming and alignment testing, with early talks under way with others, including Hugging Face.
The analogy Anthropic reaches for is telling: bank supervision. Regulators station resident examiners inside the largest banks precisely because a bank understands its own risk positions better than any outsider ever could. The same logic, the company argues, applies to a frontier AI lab. It is a striking thing for a private company to invite, and it lands in a specific context. Chief executive Dario Amodei pledged exactly this step last month, in an essay urging frontier labs to slow down after a run of agent breaches. OpenAI's Sam Altman said his company would follow suit. It has not yet. Anthropic's move is the first time any frontier developer has actually done it.
Whether embedding solves the deeper problem is another question. Earlier reporting has painted the evaluator ecosystem as thinly funded and uncomfortably dependent on the very labs it is supposed to scrutinise, and six people with building passes do not change that overnight. There is also a tension Anthropic has not resolved: an auditor given a colleague's access and a colleague's rapport may find it harder, not easier, to deliver a damning verdict. Still, after a month in which the main AI story has been models quietly breaking into systems, a company handing outsiders a real look inside is a rare piece of news that points the other way. The test will be what METR publishes, and how much Anthropic lets it sting.