Anthropic is giving employee-level access to evaluators, they are checking how safe Claude is

Anthropic is giving independent AI evaluator -- METR -- deeper access to its internal systems to assess the risks of its AI models. Six METR employees will be able to examine the company's safety practices and publish their findings without Anthropic's approval. Here's the full story.

In Short

• Anthropic is giving independent AI evaluator access to its internal systems
• The evaluator can publish its AI safety assessments without Anthropic's approval
• However, certain information remains restricted

As AI models become more capable, questions around their safety are becoming harder to answer. Anthropic, the Claude maker, has announced it will provide a research nonprofit organisation, METR, with deeper access to its internal systems. Notably, METR is dedicated to scientifically measuring whether and when advanced AI systems might pose a catastrophic threat to society.

The development follows an essay by Anthropic CEO Dario Amodei, in which he proposed that AI farms should allow third-party evaluators, or what he called "embedded evaluators", to supervise the AI labs. Later, at the White House, President Donald Trump and the leaders of top U.S. AI companies signed a voluntary agreement titled the “White House Accord on Super Intelligence.” The accord leans on industry self-policing and outlines four “layers of controls and audits.” The third layer commits companies to work with independent external auditors who can assess whether their safety controls are working as intended.

Now, Anthropic revealed they onboarded six employees from METR, giving them access to internal tools, systems and information similar to what its own risk assessment teams can access. Anthropic says METR will be able to publish its assessments of the company's risks, incidents and safety practices without requiring approval from Anthropic.

Why is Anthropic giving outsiders access

AI companies have information about their models that independent researchers and the public cannot easily access. This includes details about how models are trained, how safety measures work and what risks emerge during development.

Anthropic says it cannot publish all this information because some details could be misused by malicious actors or reveal commercially sensitive information. It is therefore giving selected external evaluators deeper access so they can examine its systems and share evidence-based conclusions.

However, this access comes with certain restrictions. METR cannot access customer data, legally privileged information or specified details about Anthropic's model architectures. The company can request that such information be removed from published reports, but METR can publicly explain whether it believes those restrictions were justified.

Anthropic says the arrangement “is limited to a small team operating under enforceable agreements.” It is not intended to give other organisations access to its internal information.

Anthropic plans three levels of independent evaluation

The company is also developing a three-tier system for its external evaluators. At the first level, evaluators conduct limited investigations, with Anthropic deciding what information they can access and approving their findings before publication. The second level offers access to more information and allows evaluators to publish findings without prior approval, subject to specific redactions.

The third level is the arrangement now in place with METR, combining employee-like access with similar publication rights.

Anthropic is also working with SecureBio, which is reviewing the chemical and biological weapons section of its August 2026 risk report. It is developing evaluation partnerships with Accenture and Faculty, while discussions with organisations such as Hugging Face are at an early stage.

- Ends