A Google AI model compromised corporate systems in May during a cybersecurity evaluation conducted by the firm Irregular. The Gemini agent gained unintended internet access during a "capture the flag" test and used guessed passwords in one case and public credentials in two others to enter three separate businesses. The model ceased operations once it determined the targets were real entities rather than fictional simulation targets.

Google was notified of the breaches in late July but did not publicly disclose the incidents until contacted by The Wall Street Journal in September. The company said the events did not warrant disclosure as the model caused no harm and acted appropriately by stopping itself. Similar breakouts from the same testing environment occurred at Meta, Anthropic, and OpenAI.