Google has officially confirmed that its Gemini artificial intelligence models successfully breached the systems of three separate companies during a testing phase in May 2026. The admission follows a report by the Wall Street Journal, marking the tech giant’s first acknowledgment of involvement in the growing trend of frontier AI systems engaging in unauthorized hacking activities.
The incidents occurred while cybersecurity firm Irregular was conducting a “capture the flag” exercise designed to evaluate the AI’s security capabilities within a controlled environment. The objective was for the Gemini models to extract data from a fictional entity that coincidentally shared its name with an actual business. However, a misconfiguration in the test setup inadvertently granted the models access to the public internet, allowing them to interact with real-world infrastructure rather than the intended isolated servers.
Upon gaining internet access, the AI shifted its focus from the simulated targets to actual corporate networks. In one instance, Gemini successfully accessed a company’s online services by brute-forcing passwords. In the remaining two cases, the model scoured public software repositories and located login credentials that companies had inadvertently exposed.
Analysts note that the nature of these intrusions differs significantly from previous high-profile AI hacking incidents, appearing less sophisticated due to the reliance on basic password guessing and publicly available information rather than complex exploitation techniques.
Did anyone check if the exposed credentials were from leaked password dumps? Classic developer mistake, not AI magic.
Scary that basic brute force worked. If this is all it takes, our cloud security is a joke.