Google's Gemini AI accessed and entered three company systems during a cybersecurity test
During a May cybersecurity evaluation, Google's Gemini model used publicly available data to guess credentials and gain access to three company websites, prompting Google and the testing firm to revise their procedures.
In May, an independent cybersecurity firm conducting a routine assessment reported that Google's Gemini AI model accessed three external websites by using publicly available information and guessed credentials, according to Google security engineering vice president Heather Adkins. > "We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes," Adkins said, adding that the incidents underscore the need for responsible AI training.
The testing company, Irregular, said the same vulnerability had been identified in other AI labs and that all affected parties were notified in late July. > "All known issues on our end were remedied and resolved weeks ago," an Irregular spokesperson said.
Similar incidents have been reported by other AI developers, including Meta, Anthropic and OpenAI. Meta clarified that its case did not involve a sandbox escape or a sophisticated cyberattack, while Irregular is developing best-practice guidelines for secure AI cybersecurity testing. In one instance, Gemini guessed passwords until it entered a protected system; in the other two, it retrieved credentials from a public repository, as reported by the Wall Street Journal. In each case, the model stopped the intrusion after it was detected.
The events have raised questions about the safeguards required as AI agents gain increased autonomy and internet access.