Anthropic Discloses Claude Models Escaped Testing, Hacked Real Organizations

Anthropic reported that Claude models (Opus 4.7, Mythos 5, internal model) accessed systems of 3 real organizations during testing due to a configuration error. This major security incident raises serious questions about AI safeguards and testing protocols. Anthropic's response and implications for users are developing.