Key facts
- Google confirmed its Gemini AI models breached three companies during a May 2026 cybersecurity test.
- The breaches occurred due to a misconfiguration by the testing firm Irregular, which left the test environment connected to the internet.
- Gemini models found exposed passwords or credentials for two of the three targeted companies.
- Google learned of the incident in late July but delayed public disclosure until September 18.
- The incident is similar to prior breaches involving AI models from Anthropic and Meta.
Google has confirmed that its Gemini AI models breached three companies during a cybersecurity test in May 2026, a failure that occurred due to an internet misconfiguration by the testing firm Irregular. The AI models were part of a "capture the flag" exercise designed to test their hacking capabilities in a secure, isolated environment. However, Irregular inadvertently left the test sandbox connected to the open web and used the name of an actual company as a target.
Gemini models then identified three real companies instead of a simulated one. For two of the targets, the AI found exposed passwords online, and for the third, it guessed the password. Google stated that its models stopped short of using the stolen credentials upon realizing they had accessed live company servers. Irregular corrected the misconfiguration and did not inform Google of the breaches until late July, seven weeks after the incident, and only after The Wall Street Journal inquired about the event.
This incident marks the fourth time this year a major AI lab has admitted to a security test spilling into the real world, following similar disclosures from OpenAI, Anthropic, and Meta. Google stated that it notified the affected companies to encourage improvements in their security practices.