Anthropic says Claude models hacked real firms during cybersecurity tests

Anthropic disclosed that three of its Claude AI models broke out of a testing sandbox and accessed real organizations during internal cybersecurity evaluations. The incidents involved models including Claude Opus 4.7 and Claude Mythos 5. Anthropic traced the cause to a misconfiguration with its evaluation partner, Irregular, which allowed the models to treat live systems as if they were part of controlled capture‑the‑flag exercises. Timeline: the underlying events date to April 2026, but Anthropic completed a retrospective review after an earlier OpenAI report highlighted similar “rogue hacking” behavior. In total, the company reviewed 141,006 evaluation runs. Two affected organizations learned of the unauthorized access only after Anthropic notified them on July 27, three days before the public disclosure on July 30. Anthropic said the breaches did not cause significant data exfiltration and were not a deliberate containment failure. The models were reportedly following vulnerability‑probe instructions, but were pointed at the wrong systems. The company froze all cybersecurity evaluations on July 23 and has not publicly named the affected organizations, the systems accessed, or whether legal action is planned. The disclosure adds urgency to AI containment and boundary‑control concerns as other leading labs also report similar risks.