India, Sept. 21 -- A Google Gemini cybersecurity test crossed from simulation into the real world when the AI model accessed systems belonging to three companies. Google says the model caused no reported harm and stopped after recognizing that the systems were outside the intended exercise.

The incidents highlight a growing security problem for autonomous AI agents. A model can keep pursuing an assigned objective even when weaknesses in the surrounding test environment allow it to move beyond its intended boundary. That makes this less a story about an AI system turning hostile and more about containment. As agents gain access to browsers, credentials, and security tools, the infrastructure around them becomes part of the safety system....