
Google main AI model, Gemini, accidentally accessed systems of three real companies during a controlled cybersecurity test.
This incident happened when an independent cybersecurity firm called Irregular was testing how AI behaves. They set up a test for Gemini, and by mistake, the internet connection was left on. Thinking some real websites were part of the test, Gemini started to explore and connect to these sites without any permission from humans.
So, how did this happen? According to Google and what they revealed about the test, Gemini used two methods to access these systems without permission. First, it guessed passwords by trying different user credentials to get past security checks on a live site. Second, it searched public online spaces for credentials that could let it access protected company resources.
Google emphasized, “The three incidents show an important point. We made sure the three companies were informed and worked with our training partner to improve their testing methods. In each case, once the AI recognized that it was interacting with a real company, its actions were stopped.”
“This incident highlights how vital it is to train powerful AI models to act responsibly,” said Heather Adkins, Vice President of Security Engineering at Google.
This event adds to a growing list of incidents involving AI technologies during testing situations. Irregular recently revealed similar unexpected behaviors with AI models from OpenAI, Meta, and Anthropic.
Unlike what happened with Gemini, Anthropic’s Claude Opus model kept trying to break in even after it suspected it was targeting a real system. Cybersecurity experts warn that mistakes may become riskier as AI models gain more freedom and better internet access.
