Google confirmed that its Gemini model, during a cybersecurity test in May 2026, gained access to and compromised the systems of three real companies, after the Wall Street Journal revealed details of the incident.

The test was conducted in collaboration with Irregular, a company specializing in testing the security of AI models, but the model's internet access remained unintentional.

In one instance, Gemini was able to guess a password and successfully access the system, while in the other two instances, it used login credentials found in a public repository.