Anthropic grants six independent METR evaluators access to its internal systems

Anthropic is hosting six METR evaluators with access to its internal systems to assess the risks of its AI models and publish their findings.

Six external researchers now have Anthropic-provided equipment, desks, and access to internal tools, under conditions close to those of its own safety teams. Members of the independent organization METR, they will be able to review the company's models, incidents, and risk management practices, and then publish their findings without prior authorization. With this system, Anthropic introduces a form of embedded evaluation, designed to give external specialists broader visibility into the inner workings of its artificial intelligence systems. Six evaluators embedded within Anthropic's teams

METR is a non-profit organization specializing in assessing the capabilities and risks of advanced artificial intelligence models.

The organization has already been collaborating with Anthropic for several months, notably through various analyses of the risk reports published by the company.

The new agreement goes further. Six METR members now have Anthropic-provided hardware, workspace in its offices, and access to internal systems comparable to those used by teams responsible for risk assessment.

Anthropic employees are encouraged to share with them the information necessary for their investigations, just as they would with colleagues working on alignment risks.

Evaluators can thus examine elements that do not necessarily appear in public reports, including evaluation methods, safety guardrails, or certain incidents observed during testing.

The system nevertheless excludes several categories of information, including customer data, attorney-client privileged communications, and certain details regarding model architecture.

Access is therefore not complete, even if it exceeds what is generally granted to external organizations. Publishing results without prior approval

One of the main elements of the agreement concerns the editorial independence of the evaluators.

METR will be able to publish its analyses on risk levels, incidents, and Anthropic's internal practices without having to obtain its approval.

The company does not have a right of veto over publication.

However, it retains the ability to require the redaction of certain information falling into pre-defined confidential categories, such as customer data, legally protected information, or certain sensitive technical details.

METR will be able to publicly note these redactions and explain whether it considers them justified, without disclosing the content concerned.

This distinction is important. Independence of publication does not mean total freedom of disclosure, but it prevents Anthropic from blocking a report as a whole solely on the grounds that it disputes its findings.

The actual scope of this provision will depend in particular on how confidentiality categories are applied in future reports. Three levels of access for external organizations

Anthropic structures its relationships with evaluators around three levels, defined according to the depth of access to internal information and publication rights.

The first level concerns targeted investigations. Anthropic determines the accessible information and must approve the publication of the results.

The second level provides broader access and allows publication without prior approval. The company retains a right to request the redaction of certain categories of confidential information.

The third level corresponds to the system set up with METR. Evaluators have access close to that of employees responsible for risk analysis, with the same publication rights as those of the second level.

Anthropic aims to gradually welcome several organizations to this third level so that they can produce their own evaluations of its models and safety decisions.

At this stage, METR is the only partner explicitly presented as benefiting from this level of integration. Independence framed by confidentiality obligations

Embedding external researchers into an AI lab's systems also raises questions of information protection.

Some evaluators may work with several competing companies. They could therefore gain access to commercially sensitive elements or information likely to be exploited outside their mission.

To limit these risks, Anthropic frames the agreement with METR with contractual commitments and internal security rules.

Only a small team from the organization benefits from the provided access. Other METR employees and researchers remain separated from this work, including when collaborating with Anthropic on other projects.

In particular, the company considers detailed information on its model architecture to be especially sensitive. Its sharing may be considered when necessary for an evaluation, but will require additional internal authorizations.

This setup aims to allow a thorough examination of risks without opening all technical and commercial information to third parties. SecureBio examines chemical and biological risks

Anthropic is working in parallel with SecureBio, an organization specializing in biosecurity issues.

SecureBio has already conducted a pilot evaluation of a previous Anthropic risk report, the conclusions of which were made public.

The organization is now examining the chapter dedicated to chemical and biological risks in the August 2026 report.

This mission currently falls under the first level of access. The information provided and the publication possibilities therefore remain more limited than in the agreement reached with METR.

An expansion of the partnership may be considered after the completion of this new evaluation. Accenture, Faculty, and Hugging Face also involved

Anthropic is also developing a collaboration with Accenture and Faculty around several evaluation activities.

The planned work includes offensive testing, or red teaming, aimed at identifying problematic behaviors, analyzing alignment mechanisms, and evaluating safety features applied to models.

The partnership also provides for the development of evaluation skills and methodologies within the organizations involved.

Anthropic also indicates