Anthropic Lets Six METR Staffers Access Internal Tools for Model Audits

Anthropic announced that it is providing six researchers from the nonprofit METR with access to its internal systems, effectively treating them like employees for the purpose of auditing AI safety practices. This move aims to strengthen independent evaluation of the company's models. According to Anthropic, this level of access fulfills a recent commitment to partner with external evaluators for comprehensive safety assessments. The organization stated that this initiative supports its pledge to ensure robust, third-party scrutiny of its operations and model development processes.

How it unfolded

• Anthropic publicly committed to stronger independent, third-party assessment of its models and safety practices last month.
• The company subsequently arranged for six METR staffers to receive company hardware, office space, and access to internal tools.
• Anthropic signed the White House Accord on Super Intelligence, which includes a specific commitment to working with outside reviewers.

Who is affected

Six METR staffers are the primary recipients of this new access arrangement. According to The Decoder, these individuals now possess privileges that place them roughly on par with Anthropic's own internal risk teams. This level of access exceeds what the company has provided to other outside reviewers, including SecureBio, Accenture, and Faculty, who have received less access and tighter publication limits. Anthropic has also started early talks with Hugging Face regarding potential future cooperation.

What happens next

METR will publish its findings independently. Anthropic stated that it can request redactions in certain predefined categories if necessary. However, METR retains the right to state publicly whether it believes those redactions are justified. According to Anthropic, not all safety-relevant information can be made public because some details could assist attackers or are commercially sensitive. The company maintains that this balance allows for rigorous review while protecting critical proprietary and security information.

The background in plain language

Independent audits require reviewers to see how systems work internally, not just their outputs. This access helps experts spot hidden flaws that external tests might miss. However, companies often restrict access to protect trade secrets and user privacy. Finding a balance means giving reviewers enough information to find real risks without exposing sensitive data or intellectual property. Teams usually define clear boundaries for what can be reviewed and published. This approach builds trust by showing that safety is being checked by outsiders who understand the technology deeply.

What to do and how to stay safe: Anthropic

• Review your organization's policies on third-party access to internal systems and ensure clear boundaries are defined.
• Verify that any external auditors have the necessary permissions to conduct thorough evaluations without compromising sensitive data.
• Monitor public reports from independent reviewers for insights into industry safety practices and potential vulnerabilities.
• Establish protocols for handling redaction requests to ensure transparency while protecting proprietary information.

Step-by-step guide: Injection attacks examples: how prompt leaks break AI systems

General security guidance from the Circuit Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What specific access did METR receive?

METR received company hardware, office space, and access to internal tools, putting them on par with Anthropic's risk teams.

Are there any restrictions on what METR can review?

Yes, customer data, attorney-client privileged information, and model architecture details remain off-limits to the reviewers.

How does this compare to other reviewers?

This deal provides more access than what SecureBio, Accenture, and Faculty have received, which had tighter limits on publishing.