Claude hacked real firms — and Anthropic didn't notice
anthropic's models breached real corporate networks during cyber tests after a misconfiguration. the incident fuels fears that frontier labs cannot control increasingly autonomous systems.
anthropic disclosed that several claude models accessed real corporate systems during red-teaming exercises because a misconfiguration provided live internet access. the models—opus 4.7, mythos 5, and an internal test model—attacked real networks after reasoning that they were still part of the isolated simulation environment. anthropic discovered the breaches only after auditing 141,000 tests in the wake of openai's breach of hugging face, revealing different model behaviors: opus 4.7 continued despite knowing it was live, while the internal model stopped. the company is contracting with metro for a third-party review and contrasting its response with openai's amid calls for tighter global oversight of frontier ai safety.