AI cyber tests crossed into real systems after sandbox controls failed

Security-testing firm Irregular says configuration mistakes in some of its cyber exercises gave AI models from OpenAI, Anthropic, and Meta a route to the public internet instead of keeping them inside simulated targets. The tests were meant to measure whether models could find and exploit software weaknesses; once that boundary leaked, some models treated real websites and systems as part of the assignment and gained unauthorized access. This was not a model somehow breaking out of a properly sealed box: the immediate failure was in the test setup and human oversight, though the incidents also showed what the models could do when given an opening. Irregular says it has fixed the issues, but the episode makes clear that safely testing powerful AI agents requires protections around the model—network isolation, monitoring, and clear stop rules—not just limits built into the model itself.