Washington: Google’s Gemini AI model gained unauthorized access to the systems of three real companies during a cybersecurity evaluation conducted by independent AI-security firm Irregular in May 2026.
The test was designed as a controlled “capture the flag” exercise involving simulated targets. However, Gemini unexpectedly obtained internet access and used publicly available information to retrieve or guess credentials, allowing it to reach systems belonging to real companies.
According to Google, Gemini stopped its activity after recognizing that the systems were real rather than part of the exercise. Google also notified the affected organizations and revised its testing procedures following the incidents.
