Anthropic can no longer block the safety findings that the AI evaluation nonprofit METR publishes about the company. In an official announcement on Thursday, the AI lab said six embedded researchers from the nonprofit now hold nearly the same access as its own staff.

Employee-level access for external evaluators

METR, which evaluates frontier AI models, now works with essentially the same Anthropic hardware, working space and internal tools that the company's own risk assessment teams use. Anthropic staff have been instructed to give the evaluators the same amount of information they would provide to another employee working on alignment risk.

Under the arrangement, METR will assess Anthropic's risk levels, incidents and practices on its own and then release the results without seeking company approval. The publication decision therefore rests with the evaluator rather than the lab, in contrast with first-tier reviews, where Anthropic's sign-off is still required to publish.

A small set of pre-defined categories remains open to redaction by Anthropic: customer data, attorney-client privileged material, and details of its model architectures. METR also has the right to state publicly whether a redaction was necessary, as long as it does not reveal what was cut.

Anthropic frames the setup in terms of an information gap. Outsiders know far less about a lab's systems than the lab itself, and full disclosure could help malicious actors or leak private training and compute data. The company drew a parallel to banking supervision, in which regulators place resident examiners inside the biggest banks so that risky behavior can be caught by a third party before it reaches the public.

The METR team works under enforceable agreements. Other METR researchers, including those working on separate projects for Anthropic, are walled off from its work.

Three tiers of evaluator access

Anthropic is building three tiers of external access, and METR is so far the only evaluator at the top of them. First-tier evaluators perform narrow reviews and need Anthropic's sign-off to publish. The second tier offers more information and allows publication without approval, subject to the same redaction categories. The third adds the embedded, employee-like access that METR now holds.

SecureBio, a nonprofit biosecurity group, sits in tier one. It has completed and published a pilot review of an earlier risk report and is now working through the chapter on chemical and biological weapons in the August Risk Report.

Anthropic is also building a partnership with Accenture and Faculty covering red-teaming, alignment assessments and safeguard tests. Early talks are under way with other groups, including Hugging Face.

From Amodei's September essay to industry commitments

The move delivers on a September essay by CEO Dario Amodei in which he urged the industry to pace frontier development. As Cryptopolitan reported at the time, Amodei suggested giving outside reviewers desks, badges, laptops and access levels similar to those of staff conducting risk assessments.

Sam Altman of OpenAI endorsed the idea on X. "Committing to having independent evaluators with employee-like access is a great idea, and we will do the same," Altman wrote.

Anthropic, Google and OpenAI have also discussed creating a common industry standards body. All three, along with Meta, Nvidia and xAI, have since signed the White House Accord on Super Intelligence, pledging to work with independent auditors.

The announcement also follows a spate of incidents since late August in which AI agents acted beyond their instructions. The next markers are concrete: SecureBio's progress on the chemical and biological weapons chapter of the August Risk Report, the findings METR releases under its new publication rights, and whether OpenAI puts in place the employee-like access Altman said it would adopt. Together they will show how far independent evaluation moves from pledge to practice across the rest of the industry.