Anthropic AI models breach three companies during tests

Anthropic, the artificial intelligence company, disclosed on Thursday, July 30, 2026, that three of its AI models gained unauthorized access to the live systems of three separate organizations during internal cybersecurity testing. The company said it discovered the incidents after launching a review of its own security evaluations following a similar disclosure by rival OpenAI on July 21.

The incidents occurred during "capture-the-flag" cybersecurity exercises conducted in partnership with the third-party evaluation firm Irregular. In each case, Anthropic's prompt instructed the Claude model that it was operating in a simulated environment with no internet access. However, due to a misconfiguration that neither company initially caught, the test environments were in fact connected to the open internet. This allowed the AI models to interact with real-world systems they believed were part of the exercise.

Anthropic reviewed 141,006 evaluation runs and identified three distinct incidents. The earliest dates back to April, meaning some activity went undetected for approximately three months. The company said it halted all cyber evaluations on July 23, identified the three...