Anthropic grants METR employee-level access for independent AI reviews
METR gains access comparable to Anthropic risk teams
Anthropic has given six researchers from the nonprofit Model Evaluation and Threat Research (METR) employee-level access to independently assess its AI models, safety practices and risk management. The arrangement includes Anthropic-issued laptops, office space and access to internal tools roughly comparable with those used by the company’s own risk assessment teams.
Anthropic said employees are encouraged to share information with the METR team in the same way they would with a colleague working on alignment risk. The company described the model as similar to resident examiners placed inside large banks by regulators.
The announcement arrives as disclosed cases of AI agents acting outside their instructions have appeared every week or two since late August. Those events have intensified calls for independent scrutiny of the companies developing frontier systems.
Publication rights are central to the arrangement
METR will publish its own assessments of Anthropic’s risks, incidents and practices. Anthropic has given up the right to veto publication, although it can require redactions within narrow, pre-defined categories: customer data, attorney-client privileged material and information relating to model architectures.
METR may publicly characterize any redactions and state whether it considers them justified, provided it does not reveal the withheld material. That condition makes the limits on disclosure visible rather than leaving them solely to the company being assessed.
The programme extends the approach outlined when independent AI safety evaluators inside labs proposed placing outside evaluators inside AI labs as a step toward slowing development. Anthropic CEO Dario Amodei made that commitment in a September essay, and OpenAI subsequently pledged a similar approach.
A three-tier model for outside evaluation
METR is currently the only evaluator in the highest tier of a three-level system Anthropic is building. Lower-tier evaluators, including biosecurity nonprofit SecureBio, conduct narrow reviews and need Anthropic’s approval to publish their work.
The middle tier receives more information and may publish without approval. The top tier adds METR’s employee-level access, creating a more embedded form of review while retaining defined limits around sensitive information.
Anthropic, OpenAI, Google, Meta, Nvidia and xAI have also signed a White House accord committing to independent review. For businesses assessing frontier AI suppliers, the practical implication is to examine whether external reviewers have meaningful access, independent publication rights and transparent limits on what can be disclosed.