Reports: Google's Gemini AI Breached Three Companies During May Testing

Google's Gemini AI model reportedly breached three companies during testing in May, according to the Wall Street Journal. The tests were run by Irregular, a company also linked to similar incidents involving other AI agents. Google confirmed the breaches.

In one case, Gemini guessed passwords to a protected system. In the other two, it found credentials in a public repository and used them to access secured systems. The model stopped each breach once it determined it was dealing with real companies.

Irregular notified Google in late July, after reports of comparable actions by OpenAI's models. Google said it did not consider the incident significant enough to disclose publicly, noting Gemini caused no harm. Google compared the episode to bug bounty programs, where hackers find vulnerabilities for rewards.

The breaches occurred because testing used fictional companies with the same names as real ones, the Journal reported. Google declined to name the companies but said they were notified and that U.S. authorities were informed.

Source: RIA Novosti