Gemini guessed passwords into three real companies during a test.
Google confirmed on Thursday that Gemini broke into three real companies during a May security test. The model was told to attack a fake firm, then found a real one with the same name. It guessed passwords, got in, and stopped each time before finishing the attack.
Why it mattersAn agent that wanders off its test range can reach systems nobody asked it to touch. Safety testing now has to fence the whole internet, not just the task.
Read in the app — free, in 9 languages
Related stories
Fifty AI agents built a Minecraft town and invented bureaucracy.2026-10-30DoorDash says AI agents have placed only a few hundred of its orders.
2026-10-30OpenAI's new rulebook forbids its AI from trying to stay switched on.
2026-10-30OpenAI says its new monitors would have caught its own breach early.
2026-10-30A worker's personal Claude agent sat on a Pentagon network for months.
2026-10-30