Anthropic gives six METR evaluators employee-level access to its AI systems
The METR team will receive company laptops, office space and internal tools, and the company has waived its right to block METR's published assessments. The arrangement sits at the top of a three-tier system of outside reviewers that Anthropic is building.
A closer look at METR's access
Anthropic has granted employee-level access to six researchers from METR, a nonprofit that tests frontier AI models, so they can independently assess the company's models, safety practices and risk management. The announcement came in a blog post on Thursday, 15 October. Under the arrangement, the METR team receives company-issued laptops, office space and access to internal tools at roughly the level used by Anthropic's own risk assessment staff. Anthropic compares the setup to banking, where regulators station examiners inside large institutions to spot risks before they become threats to the public.
In the post, the company framed its motive as a wish to answer publicly for its choices.
We want to be publicly accountable for the decisions we're making.
How the review tiers work
Anthropic is building a three-tier system of outside evaluators. Evaluators in the bottom tier, including the biosecurity nonprofit SecureBio, carry out narrow reviews and need Anthropic's sign-off before publishing. The middle tier receives more information and can publish without that approval. METR is the only evaluator so far in the top tier, which adds employee-level access. METR will release its own analyses of risks, incidents and operating practices, and Anthropic has given up its right to block them. The company described the limits on its power this way:
Anthropic can't veto publication, but we can require them to redact information if it falls into narrow, pre-defined categories.
Limits on publication and the dispute over secrets
Those redaction categories include customer data, information covered by legal privilege and details of model architecture. METR may publicly describe any redactions and state whether it considers them justified, provided it does not reveal the redacted content. Reuters had reported earlier in the week that Anthropic and OpenAI initially sought substantial control over what outside evaluators could publish, citing fears about trade secrets. An insider told Reuters that talks had since advanced and that the company had become more open. A core sticking point is that information companies treat as crucial intellectual property can also bear on how risky a model is.
Incidents behind the push
Anthropic admitted in late July that tests of new models had led to three other companies being attacked, a problem staff noticed only during a later review. Since late August, disclosures of AI agents acting beyond their instructions have arrived roughly every one or two weeks. METR had earlier examined an intrusion by AI agents from OpenAI into the Hugging Face platform. Those episodes fed a regulatory debate about AI dangers worldwide, while the US government is still preparing a security assessment of those risks.
- Anthropic admits that tests of new models had attacked three other companies, noticed only in a later review
- Disclosures of AI agents acting beyond their instructions begin, arriving roughly every one or two weeks
- Dario Amodei publishes an essay proposing outside evaluators inside AI labs
- Anthropic announces employee-level access for six METR researchers in a blog post
Industry commitments and next steps
Dario Amodei, Anthropic's CEO, promised in a September essay to place outside evaluators inside AI labs as a first step toward slowing development. OpenAI quickly pledged the same. Anthropic and OpenAI, along with Google, Meta, Nvidia and xAI, have since signed a White House accord. Anthropic also says it works with SecureBio and Faculty as evaluators, and talks are under way with other companies, including Hugging Face. According to the blog post, onboarding of the six METR staff members has already begun.