Anthropic says its own AI models breached 3 companies in security tests

Anthropic reviewed 141,006 evaluation runs after OpenAI's Hugging Face breach and found three cases where Claude reached the internet from sandboxed tests and accessed three organizations' production systems. The incidents involved Opus 4.7, Mythos 5, and an internal research model, tied to a misconfiguration with partner Irregular.