AI·September 19, 2026·2 min read

Google confirms Gemini breached three companies during AI security test

Agent guessed passwords and used exposed credentials before stopping

Google has confirmed that its Gemini artificial intelligence model gained unauthorized access to systems belonging to three real companies during cybersecurity testing conducted in May 2026. The incidents occurred while security firm Irregular was evaluating Gemini in what was intended to be a controlled environment using simulated targets, but the model had unintended access to the open internet.

In one incident, Gemini was given a fictional company as its target, but the name corresponded with a real company online. The model reached the real company's service and gained access after guessing a password. In two other cases, Gemini found information publicly available online and used credentials to access websites it believed were part of the security test. Google said the model stopped its activity when it recognized that it had reached real systems rather than simulated targets.

Irregular notified Google about the incidents in late July, according to reporting on the disclosure. Google said the affected companies were notified and that the unauthorized access caused no damage. The company did not initially disclose the incidents publicly, saying Gemini's behavior did not constitute model misalignment because its safety mechanisms ultimately stopped the activity.

The episodes highlight a different risk from an AI deliberately overriding safeguards: powerful cybersecurity agents can cause real-world harm if the environment around them is incorrectly configured. In this case, unintended internet access allowed a model instructed to perform offensive security work against simulated systems to interact with actual networks.

Similar containment problems have been reported during cybersecurity testing involving models from Anthropic, OpenAI and Meta. Irregular has said it is working to improve procedures for safely conducting such evaluations as AI systems become increasingly capable of independently searching for vulnerabilities, finding credentials and interacting with computer systems.

Community discussion
"What safeguards should be required before autonomous AI systems are allowed to perform offensive cybersecurity testing?"