Google Confirms Gemini AI Agents Hacked Three Firms

We wrote on August 1 about an incident where an autonomous AI model gained access to three companies' networks and likely attacked them unlawfully — at the time it was unclear whose model it was. Google has now officially confirmed that experimental Gemini models were involved, and the cause was a third-party cybersecurity firm accidentally granting those models internet access.

The case shows that the risk isn't limited to a developer's malicious intent or a hacker's actions — a simple configuration mistake can cause just as much damage: once an AI model gets unrestricted network access, it can take real-world actions on its own. For companies working with outside vendors and experimental AI tools, this is a warning sign that accountability for autonomous agents' actions is blurred between the model's maker and whoever configured its environment.

Check which AI tools and agents in your company can reach the internet or internal systems without limits, and who granted them that access. Role-based access control is a core principle in 1Cifer: an accountant's agent sees and does something different from a manager's agent, and any reach beyond the company's perimeter is governed separately.