Anthropic Says Embedded METR Team Can Publish Without Its Veto

Six evaluators with employee-like access

Anthropic said on October 15, 2026 that it had onboarded six employees of METR, a nonprofit that evaluates frontier AI risk, with company hardware, office space and access to internal tools and systems "roughly on par with what our own risk assessment teams have access to." Anthropic staff are encouraged to share information with them as they would with a colleague working on alignment risk. [1][2]

Three kinds of information are excluded: customer data, attorney-client privileged material and model architecture details, which Anthropic treats as a high leak risk of usually low relevance and shares only with explicit internal approval. Other METR staff, including those working with Anthropic on separate projects, are siloed from the embedded team under enforceable agreements. The post names no specific models. [1]

Publication terms and redaction limits

METR will publish its own analysis of Anthropic's risk levels, incidents and practices. "Anthropic can't veto publication," the post says, but it can require redactions in predefined categories, such as reproducing confidential intellectual property, and Anthropic decides what falls into those categories. METR reserves the right to make statements about the level of access it received. Anthropic revised a footnote on October 16 to clarify METR's current project terms. [1]

METR's first review published afterwards, on October 20, covered the misalignment section of Anthropic's August Risk Report. It said its authors were engaged before the announcement, "did not have the level of access to information described in that post," and worked under terms that gave Anthropic the right to veto publication and redact any part. METR said no redactions were made, that it changed only wording and one experiment's methodology after Anthropic's feedback, and that it expects future engagements to run on different terms. [3]

The no-veto terms had not yet produced a published report when this page was checked. The first test will be the embedded team's own analysis, and whether any redactions Anthropic requests stay inside the predefined categories. [1][3]

A ladder for other evaluators

The post sets out three tiers of outside access. In the first, evaluators run narrowly scoped risk investigations and publish with Anthropic's approval. In the second, they get more information and can publish without approval, though Anthropic can still require certain categories removed and the evaluator can say whether it thinks a redaction was justified. The third is METR's arrangement: employee-like access with second-tier publication rights. [1]

SecureBio was reviewing the chemical and biological weapons chapter of the August Risk Report under early-partner terms. Anthropic said it was in early discussions with Hugging Face and others, and working toward a broader evaluation and red-teaming partnership with Accenture and Faculty that it had first described on September 18. [1]

The September pledge behind it

In his September essay calling on labs to pace the frontier, Dario Amodei proposed that each frontier company give "ongoing, employee-like access to a team of embedded third-party evaluators" and said Anthropic would do so unilaterally, with reviewers able to publish key findings without its editorial control. On September 29 Anthropic signed theWhite House Accord on Super Intelligence, whose layers of oversight include an outside auditor. [4]

As Anthropic describes them, the October terms match the essay's promise closely: comparable access, publication without editorial control, and redaction limited to narrow categories. Anthropic said it was embedding evaluators voluntarily and that every frontier company should do the same, which makes the arrangement a benchmark that other labs' versions can be compared against. [4][1]