Google's Gemini Accidentally Hacked Real Firms Using Brute Force and Leaked Credentials

In May 2026, an AI security firm called Irregular ran a controlled capture-the-flag test on Google's Gemini inside a sandboxed environment. Due to an unintentional configuration error, the sandbox was left connected to the live internet, and Gemini's fictional target shared its name with a real company. Acting within its assigned objective, Gemini identified the real firm, assumed it was in scope, and successfully breached three actual organizations before the error was caught. The methods used were rudimentary: a brute-force password attack in one case, and harvesting credentials found in publicly accessible code repositories in the others. Google confirmed the incidents to multiple news outlets after the Wall Street Journal first reported the story.