Anthropic grants outside evaluator METR employee-level access in AI safety push (Oct 15)

(This story has been updated to remove a reference in the second-to-last paragraph of the October 15 report to Anthropic saying SecureBio had first-tier evaluator access. Anthropic had removed that detail from its blog post.)

By Aditya Soni

Oct. 15 (Reuters) - Anthropic has given the nonprofit organization METR access to its systems at a level similar to that of its employees. The company is also talking with other groups, including Hugging Face. AI companies are presenting outside reviewers as a way to make AI safer while allowing its development to continue without slowing down.

The company that develops Claude AI models is also working with SecureBio, a nonprofit focused on biological security, and Faculty, Accenture's AI business, to review its models, Anthropic said in a blog post on Thursday.

Reuters reported exclusively earlier this week that Anthropic and OpenAI had tried to keep significant control over what outside reviewers could publish. The companies were concerned that review findings could reveal trade secrets.

A person familiar with the matter told Reuters that talks with Anthropic had changed recently, with the company becoming more willing to reach an agreement.

One major disagreement between outside reviewers and AI companies is that information a company sees as important intellectual property may also be needed to judge the risks of its AI models. This can lead to arguments and to some information being removed from reports, another person familiar with the discussions told Reuters.

Experts have also raised concerns about whether outside reviewers can remain independent, since Anthropic and OpenAI choose who carries out the reviews.

The results of these talks will help determine how the world checks and monitors a technology that some of its creators warn may be developing faster than people can manage it.

THE METR PARTNERSHIP

Anthropic said on Thursday that six METR employees — METR stands for Model Evaluation and Threat Research — have company equipment, office space and access to internal systems at a level roughly equal to that of Anthropic's teams that assess risks. However, they will not be allowed to access customer data, information protected by legal privilege or the design of the company's AI models.

Anthropic said it cannot block METR's findings about risks, incidents or company practices. However, it can ask for specific, limited parts of a report to be removed. METR can also publicly challenge those requests without revealing the information that was withheld.

"In line with our transparency policies, our report will openly describe the scope, access, and other limitations of our work," METR said. On Wednesday, the nonprofit had proposed that AI developers allow private interviews with staff, publish assessments promptly, and clearly explain limits on reviewers' access and any information removed from reports.

Anthropic said its partnerships with other reviewers are at an early stage. Those reviewers currently have less access and fewer rights to publish their findings.

The company is creating three levels of access for outside reviewers. Reviewers at the first level will carry out investigations with a limited focus and must get approval before publishing their findings. Reviewers at the second level will receive more information and can publish without approval, although some information may be removed. Reviewers can publicly challenge those removals.

The third level, which METR has, gives reviewers access similar to that of employees and publication rights similar to those at the second level.

SecureBio is reviewing the section on chemical and biological weapons in Anthropic's August risk report. It has already published a trial review of the previous report. Anthropic said it would consider giving SecureBio more access after the review.

Last month, Anthropic also announced a planned partnership with Faculty. The work will cover testing AI models, testing how they respond to difficult or hostile situations, and checking whether they behave as intended, among other areas.

(Reporting by Aditya Soni and Anhata Rooprai in Bengaluru; Editing by Jonathan Ananda)