Anthropic gives METR deep access to internal tools for unhindered AI safety reviews

Anthropic has granted six evaluators from the nonprofit METR employee-like access to its internal systems and frontier models, according to ZDNET. This arrangement allows external reviewers to examine the company’s AI infrastructure with permissions similar to those held by Anthropic’s internal risk assessment teams, aiming to close the gap between what evaluators understand and what the company knows about its own systems.

Root cause

The decision stems from increasing regulatory pressure on AI labs following recent security incidents, as noted by ZDNET. Anthropic stated it wants to be accountable but cannot publicly share all safety-relevant information because some details could aid malicious actors or reveal commercially sensitive data, such as compute or training information. Providing deep access to independent parties addresses this transparency gap without exposing proprietary secrets directly to the public.

Attack path

METR evaluators are provided with Anthropic hardware and office space, enabling them to access internal tools and systems roughly on par with internal risk teams, according to the announcement. Anthropic employees are encouraged to share information with these evaluators as they would with colleagues working on alignment risk, excluding clearly defined categories such as customer data, attorney-client privileged information, or specific details about model architectures.

Affected versions

The access applies to Anthropic’s internal tools and systems that are roughly on par with what its own risk assessment teams have access to. METR, a third-party nonprofit that tests frontier models, will evaluate the risks those models pose. The company explicitly noted that researchers like METR are valuable because they do not hold equity in Anthropic, nor are they financially or operationally dependent on the company, ensuring independent oversight.

Mitigation

• METR will publish its findings, none of which Anthropic is allowed to withhold.
• Anthropic will ask METR to redact sensitive information where necessary to protect commercial interests.
• The arrangement is part of broader suggestions for “pacing the frontier” proposed by CEO Dario Amodei in a September blog post.

Why it matters

Third-party auditing introduces independent verification into AI development, reducing reliance on self-reported safety metrics. Teams usually struggle to objectivity when evaluating their own systems due to internal pressures. External reviewers provide an unbiased perspective, identifying risks that internal teams might overlook or underreport. This model helps build trust with regulators and the public by demonstrating transparency without compromising proprietary secrets. It sets a precedent for how AI labs can balance accountability with commercial confidentiality, encouraging industry-wide adoption of independent safety reviews.

What to do and how to stay safe: Anthropic

• Monitor announcements from AI labs regarding third-party audit partnerships and transparency initiatives.
• Review how your organization handles external access to sensitive systems and ensure clear data boundaries.
• Stay informed about regulatory developments in AI safety that may mandate independent testing.
• Evaluate whether your team’s risk assessment processes include independent verification mechanisms.

Step-by-step guide: Stop Sensitive Data Leaks in AI Chatbots: Priority Controls

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

Frequently asked questions

Who are the external evaluators granted access?

Six evaluators from METR, a third-party nonprofit that tests frontier models, have been given employee-like access.

Can Anthropic withhold METR’s findings?

No, Anthropic cannot withhold any findings published by METR, though it can request redaction of sensitive information.

Why did Anthropic implement this change?

Anthropic aims to close the information gap between evaluators and the company while maintaining accountability without exposing commercially sensitive data.