1. Home
2. News
3. Anthropic

Anthropic embeds six METR evaluators with employee-like access, and can't veto what they publish

Six of METR's staff now sit inside Anthropic with roughly the access its own risk teams have, and their findings go out without Anthropic's sign-off.

By Think Facility · October 15, 2026 · Anthropic AI safety Save

Anthropic said on October 15, 2026 that six people from METR, a nonprofit that evaluates frontier AI models, are now working inside the company. Their access to internal tools is "roughly on par with what our own risk assessment teams have access to."

METR will publish what it finds. Anthropic can't stop that, though it can make METR cut information that falls into a few categories agreed in advance.

Announced

Oct 15, 2026

Evaluator

METR, a nonprofit that evaluates frontier AI models

Team

Six METR employees

Publication

Anthropic can't veto publication

Other evaluators

SecureBio, Accenture and Faculty

What METR gets to see

Anthropic's staff are encouraged to share with the METR team the way they would with a colleague working on alignment risk. Customer data and privileged legal advice stay off limits. So does anything about how Anthropic's models are built.

That last one gets its own footnote. Evaluators may also work for other companies (bank examiners do too, Anthropic points out), so it's wary of things leaking by accident. Architecture details carry a high leak risk, it says, and their relevance to risk assessment is "typically low." Where they do matter, Anthropic says it'll generally share them once someone inside signs off.

If something is cut, METR can say so in public and argue it shouldn't have been, as long as it doesn't reveal what was removed. The rest of METR is walled off. Its researchers who work with Anthropic on other projects are kept out of the embedded team's work.

Three tiers, and who sits where

Anthropic is building three tiers of access. In the first, it decides what an evaluator sees, and findings go out only with its approval. In the second, evaluators see more and can publish without asking. The third adds employee-like access, and METR is the one evaluator Anthropic puts there.

SecureBio is in the first tier. It's reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, after publishing a pilot review of the report before it. Anthropic's work with Accenture and Faculty covers red-teaming models and testing their safeguards, and the post doesn't place it in a tier. Anthropic is also in early talks with Hugging Face and others.

Why a bank examiner

Anthropic compares the setup to bank supervision. Banks know more about their own risks than outsiders do, so regulators put resident examiners inside the largest ones. Anthropic says itself that the analogy "isn't exact," and the gap I'd point to is plain enough: a bank examiner answers to a regulator, and the post names no regulator behind METR.

The idea goes back to Dario Amodei's essay We Must Pace the Frontier, which named METR as an example of the evaluators he wanted inside every frontier lab. Anthropic's post also cites the White House Accord on Super Intelligence, where companies pledged to partner with outside auditors or evaluators. Its first named partner was Accenture, with Faculty.

Anthropic didn't say when METR's first report will come out.