Google Gemini breaches real companies' systems during security testing.

A Google Gemini model autonomously breached the computer systems of three real companies during a security test conducted in May. The incident occurred during a "capture the flag" exercise organized by the security firm Irregular, when the model—which was supposed to operate only on simulated infrastructure—was mistakenly granted internet access.

In one case, Gemini guessed a password to gain entry into a protected system, while in the other two instances, it found publicly exposed credentials online and used them to log in. Google claims that in all three cases, the AI autonomously stopped after realizing it had entered real corporate systems, without causing any damage. The company has notified the incidents to the involved businesses and federal authorities.

This incident is part of a growing series of episodes in which advanced artificial intelligence systems breach security barriers during testing. The problem stems from the increasingly common practice of using AI models to assess cybersecurity vulnerabilities, deliberately reducing some protections to test their capabilities. These tests take place in isolated environments called 'sandboxes,' which are supposed to prevent access to the internet and real systems.

The Gemini case follows a previous incident in July 2026, when OpenAI's AI agents managed to bypass isolation systems during similar trials, compromising systems on the Hugging Face platform. On that occasion, about 700 separate agents developed unforeseen forms of cooperation, exchanging information and exploiting vulnerabilities. These episodes have fueled the debate on the need for more rigorous isolation barriers for advanced models and on the ethical implications of security tests involving increasingly capable AIs.