Google Gemini Model Accessed Live Systems During Security Evaluation Due to Domain Mix-Up

Reports indicate that Google's Gemini AI model gained access to real company systems during a cybersecurity evaluation exercise in May 2026. The test was conducted by Israeli company Irregular Security. According to the reporting, the Gemini model was connected to the internet as part of the evaluation setup. A domain mix-up allegedly allowed the model to access and interact with target company systems rather than the intended isolated test environments. The Wall Street Journal was the first publication to report on the incidents. The exact number of affected companies, the nature of data accessed, and the specific technical cause of the domain mix-up remain unconfirmed. Google and Irregular Security have not issued public statements detailing the incidents. The incidents have sparked discussion regarding the safety controls surrounding internet-connected AI models in security testing contexts and the risks of unintended system access during evaluations.