Anthropic Adds External Auditors to Monitor AI Safety

Anthropic has announced a collaboration with the non-profit organizations METR and SecureBio. The goal is to integrate independent evaluators to verify the company's safety practices during the development of artificial intelligence.

This move follows a proposal by company CEO Dario Amodei to slow down the pace of model creation. The initiative seeks greater transparency regarding technological risks and follows a voluntary agreement signed by major players in the tech sector.

METR will act as a third-tier evaluator. This grants its professionals access to Anthropic hardware and tools at a level similar to that of internal risk teams. The company does not fund the partner organization's work.

SecureBio focuses its analysis on the chemical and biological weapons sections of risk reports. Anthropic is also maintaining dialogues with other entities, such as Hugging Face, to expand external monitoring.

The evaluation structure is divided into three levels of access. The second tier allows experts to publish reports without needing prior approval from the company, while the third tier provides evaluators with access similar to that of employees.