Anthropic Embeds Six METR Evaluators With Internal Access and No Publication Veto

Six employees of METR, the nonprofit that measures what frontier AI agents can do autonomously, now work inside Anthropic with company hardware, office space and access to internal tools that Anthropic says is roughly on par with its own risk assessment teams. Anthropic announced the arrangement today, along with the unusual part: METR can publish its analysis of Anthropic's risk levels, incidents and practices, and Anthropic cannot veto publication.

The move follows commitments Anthropic made last month to stronger third-party assessment and its signing of the White House Accord on Super Intelligence, in which labs pledged to partner with independent external evaluators.

What "employee-like access" actually grants

Outside a few pre-defined categories — customer data, attorney-client-privileged information, and model architecture details — Anthropic staff are encouraged to share information with the METR team as they would with a fellow employee working on alignment risk.

Anthropic can still require redactions in those narrow categories, but METR is free to characterize the redactions publicly and to say whether it believes they were justified. That matches METR's own framework: in a post published October 14, the nonprofit says it retains the right to attach a "redaction summary" to reports when withheld information was important to its conclusions, and commits to disclosing scope, access limits and informal pressures. It also confirms it takes no payment or donations from frontier AI companies — a point of independence Anthropic's own criteria for evaluators demand.

Three access tiers, and where each partner sits

Anthropic is formalizing a ladder rather than a single deal:

• Tier 1: narrowly scoped risk investigations; access chosen by Anthropic; findings published with company approval. SecureBio sits here, currently reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report after a published pilot review.
• Tier 2: publish without approval, broader access, redaction requests still possible, with the right to say whether redactions were justified.
• Tier 3: the METR arrangement — employee-like access plus tier-2 publication rights.

Anthropic says it wants several evaluators at tier 3 over time. It is also working with Accenture and Faculty toward partnerships covering model red-teaming and alignment assessment, and is in early discussions with other organizations including Hugging Face.

The bank-examiner logic, and its limits

Anthropic openly compares the model to bank supervision: regulators station resident examiners inside large banks because outsiders cannot see risk from published filings alone. The company says the goal is to close the gap between what evaluators understand about a frontier lab's systems and what the lab itself understands, and that every frontier AI company should embed evaluators voluntarily.

The caveats are real, and both organizations state them. Architecture information is mostly withheld because of leak risk. Other METR teams are siloed from the embedded group. METR itself warns that even with these measures it remains dependent on companies to engage, grant access and permit publication — which is why it argues voluntary arrangements still fall short of durable oversight.

What readers should watch next

The announcement is a process change, not a product one: no model, API or plan shifted. The test is the first tier-3 publication. METR has previously red-teamed Anthropic's internal agent monitoring and reviewed several risk reports under lower-access terms; an embedded analysis of incidents and practices will be the first written from inside the company's own tooling. Whether other labs copy the tier-3 terms — no veto, characterized redactions, siloed teams — is the signal to watch, alongside whether the Accenture, Faculty and Hugging Face engagements climb the tiers.

For anyone who relies on frontier models, external auditors with internal access and a duty to publish their own conclusions is a new kind of accountability primitive. Nothing in it is mandatory yet, and the first reports will show how much it actually weighs.

More on AI agent tools and workflows at LinkLoot.