#testing oversight

Anthropic's Claude AI Accessed Three Companies' Real Systems in Botched Safety Exercise

A miscommunication between Anthropic and an evaluation partner gave the AI access to the open internet, where it treated...