Anthropic gives outside safety evaluator employee-level access to its systems

Anthropic said this week that it has begun giving outside researchers a level of access to its internal systems that no AI company has previously granted a reviewer, a step the company is framing as a working model for the kind of independent oversight the industry promised regulators last month. In a post published Oct. 15, Anthropic said six employees of METR, a nonprofit that evaluates frontier AI models, have been given company hardware, office space and access to internal tools "roughly comparable" to what Anthropic's own risk-assessment staff use.

A tiered system for outside scrutiny

The arrangement sits atop a three-tier framework Anthropic has built for outside evaluators, described in the company's own announcement. Tier 1 evaluators run narrowly scoped investigations and can publish their findings only with Anthropic's sign-off. Tier 2 evaluators get more access and can publish without prior approval, though Anthropic can still require redactions in specific categories. METR's new arrangement is Tier 3, combining employee-like access with Tier 2-style publication rights: Anthropic says it cannot veto what METR publishes, only require redaction of predefined categories such as customer data, attorney-client privileged material and sensitive model-architecture details, and METR can say publicly if it thinks a redaction was unjustified.

SecureBio, a biosecurity-focused nonprofit, holds a narrower, Tier 1-style role: it is currently reviewing the chemical- and biological-weapons section of Anthropic's August risk report, after previously completing a pilot review of an earlier one. Anthropic says it will consider giving SecureBio more access once that review is done. The company is also in early talks with Hugging Face about a similar pilot.

Born of a White House pledge

Anthropic ties the move to two commitments made in the past six weeks. On Sept. 18, it announced a separate, larger partnership with Accenture's Faculty unit, under which the two companies said they each plan to invest at least $1 billion over five years to build a team of embedded evaluators who examine development and deployment decisions, not just finished models. And on Sept. 29, Anthropic was one of six frontier AI and chip companies whose chief executives signed the White House Accord on Super Intelligence, a one-page, voluntary document that calls for each signatory to partner with "an independent external auditor or evaluator" to check whether its internal safety controls are actually working. The accord sets no enforcement mechanism and leaves the definition of independence to each company.

Anthropic chief executive Dario Amodei had pushed for exactly this kind of arrangement in a September essay arguing that frontier AI companies should embed third-party evaluators with access comparable to an employee's, rather than limit them to periodic, company-controlled testing. Reporting on that essay noted that OpenAI chief executive Sam Altman said his company would commit to a similar practice, though neither company has yet matched the access level Anthropic says it has now given METR.

Who gets to call it independent

METR is not new to Anthropic's models. The nonprofit has spent the past year reviewing several of the company's periodic risk reports, including a preliminary assessment of Claude Opus 5.5's effect on AI research and development, published under what METR describes on its own site as an unpaid agreement in which Anthropic reviewed and edited METR's draft before METR signed off on the final text. METR says that review was narrowly scoped to R&D-acceleration questions and did not check compliance with Anthropic's own safety thresholds or assess the model's alignment properties.

That history is part of why the new arrangement has drawn skepticism rather than uniform praise. White House AI adviser David Sacks has publicly argued that METR cannot be treated as independent, pointing to financial and personnel overlap between the nonprofit and Anthropic's investors and staff; a separate critic has argued that METR's funding traces in part to Anthropic equity held by a major funder. Coverage of the dispute notes that METR has responded by saying it accepts no funding from AI companies or their employees and that its funders have no say over which projects it takes on; one review of the public record found no evidence that METR acts at Anthropic's direction.

"These evaluators do not reduce our accountability, but help to make it more verifiable. The safety of our models remains our responsibility."

That line, from Anthropic's own response to the independence critics, underscores the limits the company has built into the arrangement even as it publicizes it: Anthropic is choosing its own evaluators, setting the access tiers, and defining the redaction categories, even as it gives up a veto over what METR ultimately publishes within those bounds.

What happens next

An industry group calling itself the AI Evaluator Forum had already pressed for a stricter standard. In a Sept. 18 letter, it argued that embedding evaluators inside a company creates a structural conflict of interest unless the arrangement guarantees genuine independence, editorial control, transparency and protection from retaliation. Anthropic's new post does not address funding structure beyond noting that it is paying for some of this work directly in the near term, while describing the approach as voluntary and saying it believes every frontier AI company should adopt something similar.

For now, the test of the arrangement will play out in what METR, SecureBio and eventually Accenture's Faculty team choose to publish, and in whether Anthropic's redaction powers are used sparingly or expansively. Anthropic said it will post updates as the partnerships progress; neither METR nor SecureBio had published an assessment under the new access terms as of Friday.