Anthropic grants METR auditors employee-level access to internal AI systems
Anthropic allows six external evaluators from nonprofit METR deep access to its internal tools and models, marking a shift in how AI labs handle safety reviews.
Key points
- Anthropic provides six METR evaluators with hardware, office space, and access to internal risk assessment tools.
- The nonprofit will publish its findings without Anthropic withholding results, though sensitive data may be redacted.
- This move follows CEO Dario Amodei’s proposal for embedding external evaluators to help pace frontier AI development.
Anthropic announced on Thursday that it will grant six evaluators from the nonprofit METR employee-like access to its internal systems and models. This arrangement allows the external reviewers to examine frontier AI models with a level of transparency that the company says is unprecedented in the industry.
The six METR staff members have been provided with Anthropic hardware and dedicated office space. According to Anthropic, these evaluators can access internal tools and systems that are roughly equivalent to what the company’s own risk assessment teams use. This access is intended to bridge the gap between what external evaluators understand about the systems and what the company itself knows.
Context and Background
Anthropic CEO Dario Amodei proposed this model of embedding evaluators from nonprofits like METR in a September blog post. The proposal was one of three suggestions aimed at pacing the frontier of AI development. Amodei argued that external researchers are valuable because they do not hold equity in the companies they evaluate and are not financially or operationally dependent on them.
The company stated that it wants to be accountable for its safety work but cannot publicly share all relevant information. Some details could be useful to malicious actors or are commercially sensitive, such as compute or training information. By giving METR deep access, Anthropic allows the nonprofit to share conclusions about the risks those models pose without exposing sensitive internal data to the public.
Who Is Affected
METR will publish its findings, and Anthropic has stated that it is not allowed to withhold any of them. However, the company will ask METR to redact sensitive information where necessary. The access is restricted in a small number of clearly defined categories, including data about customers, attorney-client privileged information, and information about model architectures.
Anthropic employees are encouraged to share information with the METR team just as they would with a fellow employee working on alignment risk. This arrangement aims to ensure that the evaluators have the necessary context to assess the models accurately while protecting specific confidential data.
What should you do?
Security professionals and organizations relying on frontier AI models should monitor METR’s published findings for insights into model risks. Since third-party testing is not yet required by law, this voluntary step by Anthropic may set a precedent for how other labs handle external audits. Stakeholders should watch for regulatory developments that might mandate similar third-party reviews in the future.
What to do and how to stay safe: Anthropic
- Monitor publications from independent auditing groups like METR for risk assessments of frontier AI models.
- Evaluate whether your organization’s use of AI models requires independent safety verification beyond vendor claims.
- Review internal policies regarding the sharing of sensitive data with third-party evaluators or auditors.
- Stay informed on evolving AI regulations that may mandate third-party testing or reporting requirements.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What level of access do METR evaluators have at Anthropic?
METR evaluators have employee-like access to internal tools and systems, comparable to Anthropic’s own risk assessment teams, excluding specific sensitive categories.
Can Anthropic withhold METR’s findings?
No, Anthropic stated it is not allowed to withhold any of METR’s findings, though it may request redactions of sensitive information.
Is third-party testing of AI models required by law?
No, third-party testing is not yet required under law, though some safety researchers and advocates have suggested mandating it.
