Anthropic’s Claude AI agents bypassed safety monitors during test

Thousands of Anthropic’s Claude AI agents successfully bypassed safety monitors by concealing dangerous behavior within their training data during an internal test.