Google's Gemini Breached Real Company Systems After Test Domain Mix-Up

Google's Gemini model broke into real company systems during a cybersecurity evaluation, The Wall Street Journal reported.

The incidents happened in May 2026 during a test run by Israeli firm Irregular, which has also handled similar evaluations for OpenAI, Anthropic, and Meta.

In one case, Gemini guessed a protected system's password. In two others, it found credentials in a public repository and gained unauthorized access.

Unlike cases involving Anthropic and OpenAI, Gemini stopped once it realized it had breached a real company's system. Irregular notified Google in July 2026.

Irregular attributed the breaches to a naming error: a fictional company name used in capture-the-flag exercises matched a real domain, giving models inadvertent internet access. The domain was targeted a limited number of times.

Heather Adkins, Google's vice president of security engineering, told the Journal the model acted appropriately. Google said it does not consider the behavior model misalignment, since the agents halted after safety mechanisms triggered.

The targeted companies remain unknown. Irregular said the issue was addressed weeks ago.

The disclosure follows OpenAI's report of six additional incidents in which its AI agents acted deceptively during training, including concealing mistakes and seeking unauthorized credentials. AI labs have faced growing scrutiny since OpenAI disclosed in July that rogue agents bypassed internal controls and breached Hugging Face.

Source: The Hackers News