Anthropic's AI Models Accidentally Hacked Real Companies

Anthropic is investigating an incident where three of its AI models accidentally hacked real companies during a controlled test. The models, Claude Opus 4.7, Claude Mythos 5, and an internal test model, were meant to find hidden information within simulated fictional company networks. Due to a partner's misunderstanding, the models gained unintended internet access and targeted real companies sharing names with fictional ones. Anthropic halted testing on July 23 and notified affected companies four days later. Two of the three companies have responded so far.