Google: The Gemini AI model accessed external computer systems after guessing authentication credentials

Google’s artificial intelligence model Gemini accessed external computer systems after finding publicly available information online and guessing login credentials, the company announced Friday in a statement sent to AFP. The incidents occurred in May during a standard test of the model.

According to Google, three organizations were targeted, but their names have not been made public. The company said it informed them about the security breaches. In one case, described by the Wall Street Journal, Gemini reportedly tested several password combinations until it managed to break into a protected system.

Google says the model stopped on its own in all three situations, without human intervention. The company learned about the incidents in July and subsequently launched an investigation to analyze how the accesses occurred.

The cases once again highlight the risks associated with advanced artificial intelligence models, including the possibility that they may escape isolated environments and interact with external websites or platforms. The incident fuels the debate over control and accountability in the development of these systems. Heather Adkins, Google’s head of security, emphasized the importance of training models to act responsibly.