Anthropic seats six METR staff inside the company

What Anthropic says it turned on

Anthropic's October 15, 2026 post says it has started giving employee-like access to a team from METR, the nonprofit that evaluates frontier models. Six METR employees have company hardware, office space, and internal tools and systems roughly on par with Anthropic's own risk-assessment teams. Outside a short list of categories, staff are encouraged to share information with them as they would with a colleague on alignment risk. The excluded categories Anthropic names include customer data, attorney-client privileged information, and information about model architectures. An evaluation a second lab can run, from outside, is the standing briefan evaluation a second lab can run.

Publication, and the redaction limit

METR will publish its analysis of risk levels, incidents, or practices. Anthropic says it cannot veto that publication. It can require redaction when the material falls into those narrow predefined categories. METR may characterize the redactions in public and say whether it believes they were justified, as long as it does not reveal the redacted information. A right to publish with a redaction list is not the same as a joint press release. TechCrunch's account matches those terms. The artificial-intelligence desk isArtificial Intelligence.

The wall inside METR

Anthropic says the engagement is a small team under enforceable agreements. Other METR employees are siloed from that work, including people who collaborate with Anthropic on other projects. Embedded access is not a grant to the whole nonprofit. Bloomberg's brief says the same six-person shape. The privacy desk is Asked directly, Can Anthropic block a METR report? Anthropic says it cannot veto publication. It can require redaction of narrow categories it lists, including customer data, attorney-client material, and information about model architectures. METR may say publicly whether it thinks a redaction was justified, without revealing the redacted material.Privacy.

Tiers under the embedded team

Anthropic describes METR as the first evaluator at the top of a three-tier setup. It says it is also working with SecureBio and Accenture, and that it is in early discussions with other organizations, including Hugging Face. TechCrunch reports that the lower tier runs narrower reviews and needs Anthropic's approval to publish, that a middle tier gets more information and can publish without approval, and that the top tier adds the employee-level access. Where that tier map is TechCrunch's reading of the post, it stays marked as TechCrunch's. The software desk isSoftware.

What the post says it follows

Anthropic says the move follows a commitment last month to stronger third-party assessment, and that it signed a White House accord in which companies pledged to partner with independent external evaluators. This page does not restate the accord. The operational news is the six badges, the publication rule, and the redaction list. The paper is Asked directly, Is this the same as a lab buying seats? No. The Rosalind plan is a paid-seat waiver for university labs. This arrangement is a nonprofit team inside the company under agreements, with a right to publish. The access and the audience are different.Futureweb.

Who is not in the room

A university lab on a seat plan is not this team. A reader cannot apply for METR's access through Anthropic's post. The people in scope are the six named by role, Anthropic staff who are asked to brief them, and the other evaluators still at earlier tiers. The security desk is Asked directly, Who else is evaluating, on Anthropic's account? SecureBio and Accenture are named as evaluators at more limited tiers. Anthropic says it is in early discussions with Hugging Face and others. METR is the team at the top tier.Security.

What this does not decide

The post does not publish a risk rating, an incident, or a model score. It publishes an access arrangement. Whether the arrangement produces an independent account depends on what METR later prints, and on whether any redaction is explained. Until that report exists, the record is the door, not the finding. The hardware desk is Asked directly, What does Bloomberg add? A same-day account that the six-person team has hardware, office space, and internal tools, that publication is not subject to an Anthropic veto, and that redactions are limited to predefined categories. The terms this page uses come from Anthropic's post.Hardware.

How the three accounts sit

Anthropic's post is the terms. TechCrunch is a same-day unpacking, including the tier sketch. Bloomberg Law is a short confirmation of the six-person access and the veto limit. Numbers and categories in this brief come from Anthropic unless a sentence says otherwise. Energy is Asked directly, How many METR people are inside? Anthropic says six employees, on company hardware, with office space. Other METR staff are siloed from that team's work, including researchers who already work with Anthropic on other projects.Energy.

The line to keep

Six METR staff are inside, with tools comparable to the internal risk team. They can publish. Anthropic can force a narrow redaction and cannot spike the report. The rest of METR stays outside that room. The short answer this page will stand on is this: Six METR employees have company hardware, office space, and internal tools roughly comparable to Anthropic's own risk teams. METR publishes its own risk write-ups. Anthropic says it cannot veto publication, and that it can require redaction only in narrow predefined categories.