Anthropic gives METR evaluators employee-like access to check how risky its AI really is
Six staff from the nonprofit METR now work inside Anthropic with company laptops and internal tools, and Anthropic can't veto what they publish. It is the deepest outside access an AI lab has described so far.
By ScoviAI Team
Key points
- Announced October 15, 2026
- Six METR employees received company hardware, office space and access to internal systems similar to Anthropic's own risk teams
- METR will publish its risk assessments; Anthropic can require redactions in set categories but can't block publication
- SecureBio, Accenture and Faculty are also evaluating, and talks with Hugging Face are at an early stage
How do you check whether an AI company is telling the truth about the risks of its own models? Anthropic's answer is to let outsiders in, almost as if they were staff.
The company says it has started giving "employee-like access" to a team from METR, a nonprofit that tests frontier AI models. METR will publish the results of its risk assessments.
What "employee-like access" actually means
Anthropic has onboarded six METR employees. They have company hardware, office space and access to internal tools and systems that the company says are "roughly on par" with what its own risk assessment teams use. Anthropic staff are encouraged to share information with them as they would with a colleague working on alignment risk.
A few categories stay off-limits:
- data about customers,
- attorney-client privileged information,
- details of model architectures.
The key detail is who controls publication. Anthropic can't veto what METR publishes. It can require redactions only in predefined categories, such as reproducing confidential intellectual property. Anthropic decides what falls into those categories, but METR keeps the right to say publicly how much access it actually got.
The METR team works under enforceable agreements and is walled off from other METR researchers, including those working with Anthropic on separate projects.
Why a bank examiner is the comparison
Anthropic compares the setup to bank supervision. Banks know more about their own risks than outsiders do, so regulators place examiners inside the largest ones. The company argues AI labs are in the same position: some safety information can't be made public because it could help bad actors or is commercially sensitive, so trusted outsiders need to see it instead.
ADVERTISEMENT
For that to work, Anthropic says evaluators need four things: real expertise, independence (no equity in the company and no financial dependence on it), enough access, and freedom to publish.
Three tiers of access
Anthropic is building a ladder for other evaluators:
| Tier | Access | Publishing |
|---|---|---|
| 1 | Narrow investigations, access chosen by Anthropic | Only with Anthropic's approval |
| 2 | More information | Without approval; redactions limited to set categories |
| 3 | Employee-like access (the METR arrangement) | Same rights as tier 2 |
SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report. Anthropic is working with Accenture and Faculty on evaluation and red-teaming, and is in early talks with others, including Hugging Face.
Why it matters
Last month Anthropic signed the White House Accord on Super Intelligence, in which companies pledged to work with independent auditors or evaluators. This is one of the first public descriptions of what that could look like in practice. The test will be METR's published findings, and whether they say anything Anthropic would rather not hear.
Anthropic says it is embedding evaluators voluntarily and that "every frontier AI company should do the same." In Europe, outside checks are becoming a legal matter too: see what EU AI Act enforcement means for AI model makers.
Sources
- Anthropic update on independent evaluators (Anthropic)
ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.
ADVERTISEMENT
