Six METR staff now work inside Anthropic, and it cannot veto what they publish

Six staff from METR, a non-profit that evaluates frontier AI models, now have Anthropic company hardware, office space and access to its internal tools and systems, and Anthropic cannot veto what they publish about it. Anthropic says that access is "roughly on par with what our own risk assessment teams have access to", in an update published on October 15.

The limit on that freedom is redaction, not approval. "Anthropic can't veto publication, but we can require them to redact information if it falls into narrow, predefined categories," the company says, naming "data about customers, attorney-client privileged information, or information about our model architectures". METR is free to say in public that a redaction was made and whether it thinks the redaction was justified, as long as it does not reveal what was taken out.

Anthropic borrows the bank examiner model

The comparison Anthropic reaches for is bank supervision. "Banks know more about their own risk positions than outsiders do, so regulators assign resident examiners to work inside the largest banks," it says, while allowing that the analogy "isn't exact".

Inside the building the six are meant to be treated as colleagues. Anthropic says its staff are "encouraged to share information with them just as they would with a fellow employee working on alignment risk", outside the excluded categories, and that the team works "under enforceable agreements" with the rest of METR siloed from it, including METR researchers working with Anthropic on other projects.

The silo is deliberate. Embedded evaluation "is not a way for other organizations to obtain our information, or that of our customers and partners", Anthropic says.

Three tiers decide who publishes without permission

Anthropic sets out a ladder of access it is still building. A first-tier evaluator runs narrowly scoped risk investigations, with Anthropic choosing the access and approving the findings before they go out. A second-tier evaluator sees more and publishes without approval, subject to those category redactions, which it can then argue about publicly.

The third tier is what METR has: employee-like access with second-tier publication rights. Anthropic says it hopes to have several evaluators there in time, and it wants four things from each one, namely expertise that includes understanding alignment faking, independence with no equity and no financial or operational dependence, enough access to see how models were trained and how safeguards work, and the freedom to publish.

METR was reading Anthropic's risk reports before this

Anthropic says it has worked with METR on several reviews of its Risk Reports over the past year, so the relationship is not the new part. The terms are. When Anthropic announced a paid embedded-evaluation partnership with Accenture on September 18, it said only that it was in dialogue with METR about piloting elements of embedded evaluation on the evaluators' own funding, and promised to name other evaluators in the coming weeks.

The rest of that list has moved as well. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, after completing and publishing a pilot review of the previous one. That is the risk area where Anthropic loosened Claude's biology safeguards for vetted teams in September. Accenture and Faculty are "working toward" the wider partnership rather than in it, and Hugging Face is named as an organisation Anthropic is in early discussions with.

Anthropic gives no start date beyond "started", no length for the engagement and no money, and METR has published nothing from the arrangement, so there is no finding to weigh yet. Anthropic puts the rest of the industry on the spot instead: "We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same."

The first real test of the no-veto clause will be a METR report that Anthropic would rather not read.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.