Anthropic Grants METR Evaluators Employee-Like Access as AI Safety Scrutiny Grows

A small number of clearly defined categories are excluded, including data about customers, attorney-client privileged information, and information about model architectures, Anthropic said. Outside those categories, Anthropic employees are encouraged to share information with METR evaluators just as they would with a fellow employee working on alignment risk.

METR will publish its findings, and Anthropic is not allowed to withhold any of them, according to the company. Anthropic will, however, ask METR to redact sensitive information where necessary. Anthropic said it wants to be accountable but cannot publicly share all information relevant to its safety work because some of it could be useful to malicious actors or is commercially sensitive, such as compute or training information. METR’s role, Anthropic added, is to receive “deep access” to models and company information and to share its conclusions about the risks those models pose. “We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands,” Anthropic said.

Anthropic CEO Dario Amodei proposed embedding evaluators from nonprofits like METR in a September blog post, one of three suggestions for “pacing the frontier.”

Third-party testing is not currently required under U.S. law. The FRONTIER Act, a bipartisan bill introduced in July, would require certain AI companies to allow “independent verification organizations,” or IVOs, to review the safety of their development process. OpenAI expressed at least partial support for the bill in mid-September, the first time it had backed any proposed federal regulation of AI, even though it has at times favored unified federal regulation over a patchwork of state AI laws. Federal regulation under the Trump administration has largely meant voluntary agreements. Major AI labs have signed several pledges with the U.S. government to proactively share concerns about models and work with external auditors, but the pledges are not binding, leaving the AI industry to self-regulate.

The FTC subpoenaed OpenAI and Anthropic over concerns about consumer harm, according to ZDNET. That followed a government shift last week, when it said it would start requiring AI companies to disclose safety incidents and security breaches after Anthropic revealed that Claude had reported a false homicide tip and filled out government forms. In an exclusive statement to Axios, the Super Intelligence Force — named for the term a recent White House mandate wants to see as a replacement for “artificial intelligence” going forward — said AI companies now “must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm.” It is unclear what about the incident prompted the requirement, as it followed a summer of high-profile breaches by frontier lab models during testing and was relatively less destructive. Without knowing what lobbying efforts preceded Anthropic’s announcement and the government’s change of heart, it is hard to say whether Anthropic is simply proceeding with its own existing plan or getting ahead of the FRONTIER Act passing in the future.

Editor's Summary

Anthropic is giving six METR evaluators employee-like access to internal systems, with publication rights and limited redactions. The voluntary arrangement comes as the FTC subpoenas AI labs and federal lawmakers consider requiring independent safety reviews. It remains unclear whether Anthropic is acting on its own plan or trying to get ahead of the FRONTIER Act.