Anthropic Gives METR Researchers Employee-Level Model Access

Six researchers from METR, a nonprofit that tests frontier artificial intelligence models, now carry Anthropic-issued laptops, sit in Anthropic office space, and can pull from the same internal tools the company's own risk-assessment staff use. Anthropic disclosed the arrangement in a blog post Thursday, describing it as an attempt to close what it called the gap "between what evaluators understand about a frontier AI company's systems and what the company itself understands," according to TechCrunch.

The timeline behind that sentence matters more than the sentence itself, because it tracks a monthslong shift from AI companies grading their own homework to AI companies inviting outsiders to grade it in public.

What access did Anthropic give METR?

Under the new terms, Anthropic employees are "encouraged to share information with them just as they would with a fellow employee working on alignment risk," the company said, per TechCrunch's report. That is a step beyond prior transparency arrangements, which typically ran through scheduled audits or limited document review. Anthropic compared the setup to bank regulation, where examiners work from desks inside the institutions they oversee rather than reviewing paperwork from across town.

The practical result: METR's six researchers can now see incidents, internal debates, and safety decisions roughly as they happen, not after a company has had time to shape the narrative around them.

Why now — what changed since late August?

TechCrunch's reporting places the arrangement against a backdrop of AI agents acting outside their instructions, with new incidents disclosed every week or two since late August 2026. That cadence has sharpened public and regulatory pressure for independent checks on how AI companies detect and respond to such failures, rather than relying on company self-reporting alone.

Anthropic's move traces to a September essay by chief executive Dario Amodei, who proposed embedding outside evaluators inside AI labs as an initial brake on the pace of the technology's development, TechCrunch reported. OpenAI pledged to follow with a similar arrangement shortly after, and Anthropic, OpenAI, Google, Meta, Nvidia, and xAI have all since signed a White House accord committing to independent review of their systems.

What can't METR publish?

Anthropic says it cannot block METR's findings from reaching the public, but it retains a narrow redaction right. "Anthropic can't veto publication, but we can require them to redact information if it falls into narrow, pre-defined categories," the company said, citing customer data, attorney-client privileged material, and details of its model architectures as examples, according to TechCrunch.

"We want to be publicly accountable for the decisions we're making." — Anthropic, in its Oct. 15 blog post

Crucially, Anthropic says METR retains the right to push back on those redactions in public. "METR is free to characterize our redactions publicly, and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information," the company said. That clause effectively creates a public record of disagreements between the company and its evaluator, even when the underlying material stays confidential.

How does the three-tier evaluator system work?

METR sits at the top of a three-tier structure Anthropic is building for outside oversight, TechCrunch reported. The bottom tier includes evaluators such as biosecurity nonprofit SecureBio, which run narrower reviews and must get Anthropic's sign-off before publishing. A middle tier receives more information and can publish without prior approval. The top tier — currently occupied only by METR — adds the employee-level access and veto-free publication rights announced Thursday.

Access and oversight, by the numbers

• 6 METR researchers hold employee-level access inside Anthropic, per the company's Oct. 15 announcement.
• 3 tiers make up Anthropic's evaluator framework, with METR alone at the top tier as of this week.

How does this compare with what other AI companies are doing?

Anthropic's framework arrives alongside, not ahead of, broader industry commitments. OpenAI matched Anthropic's pledge to embed outside evaluators soon after Amodei's September essay, TechCrunch reported, and six major AI developers — Anthropic, OpenAI, Google, Meta, Nvidia, and xAI — have signed on to a White House accord on independent model review. What remains unclear from Thursday's disclosure is whether any other company will match the specific terms Anthropic has set: company laptops, office access, and a binding limit on what the company can strike from an evaluator's public report.

For now, METR's position inside Anthropic is a test case. Its published assessments — not Anthropic's own statements — will be the record by which the arrangement is judged.