Anthropic Grants METR Employee-Level Access for Independent Model Safety Assessments

Anthropic has announced it will provide METR with employee-like access to its infrastructure. The company stated this step is part of its broader pledge to partner with external evaluators for assessment of its artificial intelligence models and safety protocols.

According to a report by Techmeme, Anthropic confirmed that it is giving METR employee-like access. The company linked this action to a commitment made last month to pursue stronger independent, third-party assessment of its models and safety practices. This arrangement allows METR to partner with Anthropic for the assessment of its models and safety practices.

Context of the Safety Pledge

Anthropic explicitly linked this access arrangement to its public commitment made last month regarding independent oversight. The company stated that it pledged to pursue stronger third-party assessment of its models and safety practices. By granting METR employee-like access, Anthropic is operationalizing that promise, allowing an external partner to examine the inner workings of its systems with a level of privilege typically reserved for internal staff.

This development reflects Anthropic’s specific pledge to partner with evaluators. Instead of relying solely on internal audits, Anthropic is enabling a partner to interact with its systems as if they were an employee. This approach allows for the assessment of models and safety practices as described in the company’s recent announcement.

Who Is Affected

The primary entities affected by this announcement are Anthropic and METR. METR, acting as the external evaluator, gains new capabilities to inspect and test Anthropic’s models. For Anthropic, this represents a significant step in its strategy to demonstrate accountability and safety through external validation. The broader AI safety community may also view this as a potential model for how other companies might engage with independent auditors to verify their safety claims.

This arrangement does not directly impact end-users of Anthropic’s products in the short term, but it may influence the reliability and safety perceptions of those products over time. Security professionals and researchers interested in AI safety will likely monitor how this partnership evolves and what findings METR publishes as a result of its enhanced access.

What should you do?

Security and technology professionals should monitor official channels from Anthropic and METR for updates on the scope of this access. Organizations using Anthropic’s models may want to review their own vendor risk management processes to understand how third-party assessments impact service reliability. It is also advisable to stay informed about emerging standards for independent AI safety evaluation, as this partnership may set precedents for industry-wide practices.

What to do and how to stay safe: Anthropic

• Monitor official announcements from Anthropic and METR for details on the specific scope of the granted access.
• Review current vendor risk management protocols to assess how third-party safety evaluations impact your organization’s use of AI services.
• Stay informed about developing industry standards for independent AI safety audits to anticipate future compliance or security requirements.
• Evaluate whether your organization requires similar independent verification for the AI models you deploy in critical systems.

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 kind of access is Anthropic giving to METR?

Anthropic is providing METR with employee-like access to its systems to facilitate safety assessments of its AI models.

When did Anthropic make the commitment that led to this access?

Anthropic made the commitment to stronger independent third-party assessment last month, according to the report.

Why is Anthropic partnering with METR for this assessment?

The partnership is part of Anthropic’s pledge to ensure independent evaluation of its models and safety practices.