Google's Gemini AI Accidentally Accesses Three Real Companies

Google's Gemini AI system accessed three real companies in an unexpected incident during testing. The company was conducting a simulation to test Gemini's cybersecurity capabilities, but internet access was inadvertently enabled, allowing the AI to interact with external resources.

The model used simple methods to gain access to the protected systems, including trying passwords until it found one that worked and discovering credentials in public repositories. In all three cases, Gemini stopped its actions after accessing real resources.

Google and Irregular, the independent company conducting the testing, responded promptly to the incident by strengthening controls around the test environment and creating a dedicated team to assess model isolation and control mechanisms.