Anthropic Becomes One of First to Allow Third Party Test of AI Dangers

Anthropic has granted six researchers from AI evaluation nonprofit METR access to independently assess its AI models, safety practices, and risk management. This move follows a series of incidents reported since late August that have heightened concerns about AI agents operating beyond their instructions. Anthropic aims to increase public accountability and narrow the gap between the company's knowledge of its systems and that of external evaluators.

METR researchers have been provided with equipment, office space, and access to tools similar to those used by Anthropic's own risk assessment teams. Employees are encouraged to share information with the researchers in the same manner as internal colleagues. Anthropic cannot prevent METR from publishing its assessments of risks, incidents, and safety practices.

However, the company can require METR to redact specific information, such as customer data and model architecture details. METR is the sole evaluator at the highest tier of Anthropic's three-tier independent evaluation system, which also includes two lower tiers with varying levels of access and publication rights. This arrangement follows Anthropic CEO Dario Amodei's proposal in a September essay to embed outside evaluators directly within AI labs, a step towards addressing frontier AI development risks.

Anthropic, OpenAI, Google, Meta, Nvidia, and xAI have collectively signed a White House agreement to support independent review of their AI systems.