Anthropic is giving a team of independent AI evaluators unprecedented access to its internal systems, allowing them to scrutinize model risks and safety practices with fewer restrictions on what they can publicly disclose.

The arrangement with nonprofit evaluator METR is designed to give outside researchers a closer look at how Anthropic assesses potential risks, including access to information that would otherwise be difficult to verify from the outside, the company revealed in a blog post.

Anthropic said it has onboarded six METR staff members, providing them with company devices, workspace access, and entry to internal tools and systems. 

METR will be able to publish findings on risk assessments, incidents, and safety practices without Anthropic having the power to block publication, the company said. However, Anthropic can request limited redactions for information falling into predefined categories, including customer data, attorney-client privileged material, and sensitive model architecture details.

The arrangement represents the most extensive of the three levels of evaluator access, ranging from narrowly scoped reviews approved by the company to embedded arrangements that allow evaluators to work within its internal systems.

Anthropic is also working with SecureBio, Accenture and Faculty, with preliminary discussions underway with other organizations, including Hugging Face, the company noted.

SecureBio currently operates under the framework’s first tier. The organization is reviewing the chemical and biological weapons section of Anthropic’s August risk report after completing a pilot review of an earlier report. Anthropic said it will consider expanding SecureBio’s access after the current assessment concludes.

The announcement comes as AI developers face growing questions about whether their internal safety evaluations provide sufficient independent scrutiny of increasingly capable models. 

Earlier this week, the Federal Trade Commission issued civil subpoenas to OpenAI and Anthropic as regulators intensified scrutiny of AI safety incidents and potential consumer protection violations.

Meanwhile, New York City Comptroller Mark Levine is urging Anthropic to pause its planned initial public offering (IPO) until the artificial intelligence company addresses concerns about AI safety, corporate governance, and the risks of rapidly advancing technology.

Anthropic said its approach is intended to address the information gap between what outside researchers can observe and what companies know about their own systems.

The company also pointed to its signing of the White House Accord on Super Intelligence, which included commitments to work with independent external evaluators.

Anthropic compared the proposed oversight model to supervision in the banking industry, where outside authorities can gain access to internal operations to assess risks that may not be apparent through public disclosures alone.

"For the AI industry, this is a step toward robust safety oversight. We’re moving quickly with multiple evaluators and learning actively about what kinds of partnerships work. We’re committed to sharing updates as we go," the company stated.

Photo: Shutterstock