In May 2026, a Google Gemini model gained unauthorized access to three real-world corporate systems during a cybersecurity test. The experiment was conducted by the firm Irregular to evaluate AI capabilities in identifying vulnerabilities within simulated environments.
The breach occurred due to a configuration error that allowed the AI to access the public internet. In one instance, the model matched a fictional company name with a real entity, successfully guessing a password to gain access. In two other cases, Gemini utilized credentials discovered in public repositories to log into corporate systems.
Google stated that the model ceased activity upon realizing it was interacting with real organizations rather than test infrastructure. The company reported no damage was caused by these incidents. Affected organizations were notified after the breach was discovered, though their identities remain undisclosed.
The events were not publicly disclosed until inquiries were made by The Wall Street Journal. Similar unauthorized access incidents have been previously documented during the testing of AI systems developed by OpenAI and Anthropic.
