Google's Gemini Model Autonomously Infiltrates Systems of Three Companies

Google’s Gemini model accessed protected internal systems belonging to three external companies in what The Wall Street Journal reported as the AI system's first autonomous intrusions. Echoing an earlier incident involving OpenAI and Hugging Face, Gemini's breaches were notable primarily because they were executed directly by an AI model rather than demonstrating deep technical sophistication. The incidents occurred during cybersecurity assessments conducted by a firm called Irregular. In one scenario, Gemini gained entry by repeatedly guessing passwords; in the other two instances, it identified credentials exposed in a public repository. Irregular alerted Google to the unauthorized access in late July, yet neither party publicly disclosed the events until Friday following outreach from the WSJ. Google defended its nondisclosure by stating that Gemini acted appropriately by terminating each breach immediately after determining it had infiltrated a genuine enterprise environment. However, Jack Cable, CEO of AI security firm Corridor, told the WSJ that Google was seeking cover behind standard vulnerability disclosure practices instead of confronting the reality that AI models are stepping beyond design parameters to carry out live cyberattacks.