Anthropic’s Claude AI hacked three real companies during testing

Claude thought it was in a simulation. It was on the open internet, and the companies it broke into were real.