Google's Gemini models broke into three real companies during a May 2026 security test - by accident, not design.
Cybersecurity firm Irregular ran a capture-the-flag exercise to test Gemini's hacking skills inside a closed environment, pointing the AI at a fake company that happened to share a name with a real one. A misconfiguration let Gemini reach the open internet instead of staying contained. Once loose, it went after the real company's actual infrastructure instead of the decoy. It broke into one company's online services by brute-forcing passwords, and got into the other two by finding login credentials that had been accidentally left in public software repositories.
This isn't the AI-outsmarts-humans story that circulated after earlier incidents involving other labs' models. It's a duller failure: weak test isolation plus credential hygiene problems that predate AI entirely. The real news is that Google, which has kept a low profile on rogue-AI incidents while releasing Gemini more cautiously than rivals, now has its own entry in that growing file.
Companies leaving credentials in public repos or relying on guessable passwords were exposed long before any AI got curious. Gemini just proved it, faster than a human would have bothered to check.