Publication date: 16 October 2026
Anthropic has embedded six researchers from the nonprofit Model Evaluation and Threat Research group, known as METR, inside the company with access roughly comparable to its own risk-assessment teams. The arrangement is one of the most concrete attempts yet to solve a central problem in frontier AI oversight: outsiders cannot evaluate risks they are not allowed to see.
The METR team receives company hardware, office space and access to internal tools. Anthropic says employees may share information with the evaluators much as they would with internal colleagues working on alignment risk, except for defined categories such as customer data, legally privileged material and some model-architecture information.
Most importantly, METR will be able to publish its assessment without an Anthropic veto. The company can require narrow redactions, but METR may publicly characterise those redactions and state whether it considers them justified.
Access has been the missing ingredient
External evaluations often test a released model through an interface. That can reveal capabilities and obvious safeguards, but it provides limited visibility into training decisions, internal incidents, monitoring systems and the way a company responds when risks are discovered.
Anthropic’s model borrows from resident supervision in banking, where examiners work inside institutions because management inevitably knows more about its own risk position than the public does. The analogy is imperfect: METR is not a statutory regulator and cannot impose penalties. But embedded access can narrow the information gap that makes purely external scrutiny weak.
The arrangement follows growing debate over whether AI companies were seeking too much control over what outside reviewers could publish. Reuters reported that intellectual-property concerns and publication rights had become sticking points in negotiations between laboratories and evaluators. Anthropic’s no-veto commitment for METR is therefore a meaningful governance change, even with redaction safeguards.
Independence is more than permission to publish
An evaluator can have deep access and still face structural pressures. The company selects who enters, defines excluded information and controls the security environment. A small embedded team may also become socially or operationally dependent on the organisation it is assessing.
Credibility will depend on how METR is funded, whether it can choose its own lines of inquiry, how disagreements are disclosed and whether other qualified evaluators receive comparable access. Anthropic says the six researchers operate under enforceable agreements and are separated from other METR teams, including researchers working with Anthropic on different projects.
The company is developing three access tiers. Narrow first-tier reviews require Anthropic approval before publication. Second-tier evaluators can publish without approval and receive more information, while third-tier arrangements such as METR’s add employee-like access. Anthropic is also working with SecureBio, Accenture’s AI business Faculty and discussing partnerships with other organisations.
A voluntary model needs external verification
The initiative could establish a useful template for AI assurance: qualified teams, meaningful internal access, protected publication rights and transparent handling of redactions. It may also help governments distinguish substantive auditing from superficial red-teaming exercises designed mainly for public relations.
But voluntary oversight cannot answer every public-policy question. Companies can withdraw access, choose favourable evaluators or limit the consequences of adverse findings. Regulators may still need minimum standards covering evaluator competence, conflicts of interest, incident reporting and protection against retaliation.
The real test will come when METR identifies something Anthropic would prefer not to disclose. If the evaluator can publish a clear, evidence-based criticism and the company responds without weakening the arrangement, embedded assessment will gain legitimacy. Access creates the possibility of genuine scrutiny; independence will be proven only when that scrutiny becomes uncomfortable.