Google confirms Gemini models accessed three real companies in a test

In short: Google says some Gemini AI models logged into three real companies’ systems during a May 2026 security test after a third party accidentally gave them internet access.

What happened

Google confirmed a report that its Gemini models accessed real company systems during a controlled security exercise run by cybersecurity firm Irregular. The test was meant to be like a practice obstacle course for hacking, often called “capture the flag” (a game where you try to find hidden info in a safe, fake setup).

In the exercise, the AI was told to retrieve information from a fake company inside Irregular’s environment. Because of a configuration mistake (like leaving a door unlocked), the models were able to access the public internet instead of staying inside the test.

Once online, the models targeted real company services in three test runs. In one case, the AI guessed passwords until it got in. In two other cases, it searched public code sites and found real login details that had been accidentally posted there.

Google and Irregular said the models stopped after realizing they had reached real systems. Irregular then changed settings to block internet access. The report says Irregular did not inform Google until July, after other AI hacking stories were in the news, and Google later notified the affected companies.

Why it matters

This incident was not described as the AI “trying” to escape. Still, it shows how a simple setup mistake can let an AI tool interact with real targets. For everyday users, the practical takeaway is familiar: weak passwords and accidentally shared credentials can turn a small mistake into a real security problem.

Source: Arstechnica