← WORLD Δ home

METR evaluators reportedly gain employee-like access to Anthropic systems

AI securityUnverifiedUpdated:

Six METR evaluators have reportedly been embedded at Anthropic with access roughly comparable to internal risk teams, while retaining the ability to publish analyses without company approval subject to specified redactions.

METR evaluators reportedly gain employee-like access to Anthropic systems

Internal tools provided to six outside evaluators

According to the report, Anthropic gave six employees of the nonprofit evaluator METR company hardware, office space, and access to internal tools and systems.

Anthropic reportedly encouraged staff to share information with them as they would with a colleague working in risk assessment. METR’s evaluators can publish their analysis, although Anthropic may require redactions for specified categories including customer data, attorney-client privileged material, and model-architecture information.

The report does not describe a comparable prior access arrangement for METR at Anthropic.

Anthropic reportedly onboarded six METR employees, granting access to hardware, internal tools, and systems roughly comparable to that of its own risk-assessment teams. METR can publish its analysis without Anthropic approval, subject to specified redactions.

Publication rights and specified redaction categories

The reported arrangement would place an outside evaluator inside an AI developer’s operations with access close to that of internal risk teams while preserving a route for publication outside the company’s approval.

Sources

  • Ars TechnicaAnthropic gives outside AI watchdog “employee-like” access
    Read source →