Anthropic gives METR researchers employee-level access for AI reviews

In short: Anthropic says it will give six outside researchers from METR employee-level access to review its AI models and safety work, and METR can publish findings without Anthropic’s approval.

What happened

Anthropic, the company behind the Claude AI models, said it is embedding six researchers from METR inside the company to independently evaluate Anthropic’s systems. METR is a nonprofit that tests advanced AI models, including how they might fail or be misused.

Anthropic said the METR team will get company laptops, office space, and access to internal tools similar to what Anthropic’s own risk teams use. In simple terms, Anthropic is letting outsiders look “under the hood” (like a mechanic being allowed to inspect the engine, not just take the car for a short drive).

METR will publish its own assessments of Anthropic’s risks, incidents, and practices. Anthropic says it is giving up the right to block those reports. Anthropic can only ask for limited redactions, like removing customer data, legal communications, or certain details about how its models are built.

Anthropic compared the setup to banking regulators placing examiners inside big banks. The company also described this as the top level of a new three-tier evaluator system, where lower tiers get less access and sometimes need approval to publish.

Why it matters

AI systems are increasingly being used as “agents,” meaning they can take actions in software, not just answer questions (like an assistant that can click buttons for you). Recent reports of agents behaving in unexpected ways have increased pressure for independent oversight. Letting outside reviewers publish without company sign-off could make it easier for the public to understand how AI companies handle safety problems.

Source: TechCrunch AI