Gemini hacked into the protected systems of three real companies during testing - media

Google's artificial intelligence (AI) model Gemini, during testing in the cybersecurity field, independently gained access to the protected systems of three real companies.

This is reported by The Wall Street Journal (WSJ).

According to the newspaper, in one case the model tried passwords until it gained access to a system. In the other two cases, Gemini found credentials in open sources and used them to penetrate company systems. In all cases, the model stopped after realizing it was not working with a test environment.

The incidents occurred in May during a Capture the Flag exercise. A fictitious company in the test environment had the same name as a real one. Google stated that it does not consider this to be uncontrolled behavior by the model, as built-in safety measures caused it to stop. The company notified the affected companies and authorities but did not disclose the information publicly.