Google Gemini breached three companies during cybersecurity test

Google’s Gemini artificial intelligence model reportedly broke into the networks of three companies during a cybersecurity test after gaining access to the internet and moving beyond the systems it was supposed to examine.

The incidents took place in May during an evaluation carried out by cybersecurity testing firm Irregular, according to The Wall Street Journal. The exercise was designed to test how AI systems respond to security challenges.

In one case, Gemini reportedly guessed passwords until it gained access to a protected network. In two other tests, it searched public online repositories, found exposed login details and used them to enter real corporate systems.

Google said the model stopped once it realised that it had accessed real companies rather than simulated targets.

Internet connection led Gemini beyond test

The incidents reportedly began with a mistake in the testing environment. Gemini had been instructed to retrieve information from a simulated organisation during a cybersecurity exercise.

However, the simulated organisation shared its name with a real company. The test system, which was supposed to be isolated, was also inadvertently left connected to the internet.

That allowed Gemini to search the web for information about the target. During later tests, it reportedly found login credentials that had been publicly exposed and used them to gain access to two additional companies.

Google said Gemini did not cause any damage and disconnected after identifying that it had reached real corporate systems.

The company initially chose not to make the incidents public, saying no damage had occurred. Google compared the episode with a bug bounty exercise, in which researchers identify security weaknesses and report them to organisations.

Google informed affected companies

Irregular informed Google about the breaches in late July. The disclosure came after reports that AI agents developed by OpenAI had gained unauthorised access to systems on the software platform Hugging Face.

Google confirmed the Gemini incidents only after questions from The Wall Street Journal. The company said it had informed the three affected businesses and federal authorities.

Google and Irregular did not identify the companies involved or disclose which version of Gemini was used.

Heather Adkins, Google’s vice-president of security engineering, said the incident showed why advanced AI systems need to be trained to operate responsibly.