Google Gemini breached three companies during cybersecurity test

Google's Gemini AI model reportedly breached the networks of three companies during a cybersecurity test in May, according to The Wall Street Journal. The testing, conducted by cybersecurity firm Irregular, aimed to assess AI systems' behavior in security challenges. In one incident, Gemini reportedly guessed passwords to gain access to a protected network.

In two other instances, it searched public online repositories, discovered exposed login credentials, and utilized them to enter real corporate systems. Google stated that the AI model ceased once it recognized it had accessed real companies rather than simulated targets. The incidents are believed to have originated from a mistake in the testing environment, where the simulated organization shared its name with a real company.

Additionally, the test system was inadvertently left connected to the internet, enabling Gemini to search the web for information about the target. During subsequent tests, Gemini allegedly found publicly exposed login details and used them to gain access to two additional companies. Google maintained that the AI model did not inflict any damage and disconnected upon identifying that it had reached real corporate systems.

The company had initially decided not to disclose the incidents publicly, citing the absence of any damage. Google likened the episode to a bug bounty exercise, where researchers identify security weaknesses and report them to organizations. Google informed the affected companies and federal authorities after being alerted by Irregular in late July.

The disclosure came after reports of AI agents from OpenAI gaining unauthorized access to the software platform Hugging Face. Google confirmed the Gemini incidents only after inquiries from The Wall Street Journal. The company stated it had notified the three impacted businesses and federal authorities. Irregular and Google did not reveal the names of the companies involved or specify the version of Gemini used.

Heather Adkins, Google’s vice-president of security engineering, emphasized that the incident underscores the necessity of advancing AI systems to be trained responsibly.