Anthropic Embeds Six METR Staff With Employee-Like Access and No Veto Over Publication

Anthropic said on Oct. 15 that a team of six employees from METR, a nonprofit that evaluates frontier AI models, now has "employee-like access" to the company, including hardware, office space and internal tools roughly on par with its own risk assessment teams, according to Anthropic's post.

METR will publish its risk assessments, and Anthropic cannot veto publication. The company can require redactions only in narrow, predefined categories: customer data, attorney-client privileged information and model architecture information. METR can say publicly whether it thinks a redaction was justified, without revealing what was removed. Other METR staff are walled off from this team's work.

The post names no individuals. It does not say who pays for the METR team.

Three access tiers, one embedded team

Anthropic describes three tiers of evaluator access. METR sits in the third.

Tier	Access	Publication
1	Narrowly scoped risk investigations	Anthropic approves what is published
2	Broader access	Published without Anthropic's approval, with removal of specific categories allowed
3	Employee-like access	Same rights as tier 2 (the METR arrangement)

Anthropic wrote that evaluators need expertise, independence from equity or financial and operational ties, enough access and the freedom to publish. It compares the model to bank supervisors who place resident examiners inside large banks. "We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same," the post says.

Who else is on the list

SecureBio is in the first tier. It is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, after publishing a pilot review of an earlier report. METR reviewed several of Anthropic's Risk Reports over the past year.

Accenture and Faculty are working toward a broader partnership covering model evaluation, red-teaming, alignment assessments and safeguard testing. Hugging Face is in early discussions.

The post says Anthropic signed the White House Accord on Super Intelligence last month, in which companies pledged to work with independent external auditors or evaluators.

The Accenture question

The Next Web reported that on Sept. 18 Anthropic named Accenture as its first embedded evaluator, with the work led by Faculty, a British AI company Accenture acquired in January. Each company making the commitment expects to invest at least $1 billion over five years, according to the report. Anthropic pays for the work directly and says that is not the right long-term model, preferring pooled or government funding that does not yet exist.

The same report notes that Accenture is Anthropic's largest Claude Code deployment, with about 30,000 staff being trained on Claude. It argues that the evaluator's invoice still goes to the company being evaluated. Anthropic's Oct. 15 post does not address that criticism directly. The Next Web reported that Anthropic was also talking with METR and other nonprofit evaluators about piloting parts of embedded evaluation on their own funding.

Related coverage on this site includes Iowa's lawsuit against OpenAI over ChatGPT safety and the iFixAI agent auditor.

Anthropic's post does not say when METR's first assessment from the embedded team will appear, and it gives no date for the tier 2 arrangements it describes.