In the first recorded instance of Google’s AI breaking out, Gemini hacked three businesses.

During a test of its cybersecurity capabilities, Google’s Gemini model gained access to the internet and hacked other businesses; this is the first instance of the company’s AI systems carrying out such an act on their own. The intrusions happened in May when a cybersecurity test was being carried out by an independent cybersecurity assessment business called Irregular.

Heather Adkins, Google’s vice president of security engineering, said in a statement that during a routine testing evaluation, Gemini discovered public information online and guessed credentials to access three websites it believed were within the test’s scope.

“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.“These ‌events highlight the importance of training powerful AI models to act responsibly.”

An Irregular spokesperson said the incident involved the same issue that affected other AI labs and that all relevant labs were notified in late July “All known issues on our end were remedied and resolved weeks ago,” the spokesperson said.

Meta, Anthropic, and OpenAI reported similar incidents connected to Irregular. Irregular stated that it was developing best practices for safely carrying out AI cybersecurity assessments, while Meta stated in August that the event did not include a sandbox escape or sophisticated cyberattack.

The events have prompted concerns about the security measures required as AI agents become more independent and have more access to computers and the internet. In one instance, the Gemini model used password guessing to get into a secure system. According to the Wall Street Journal, which broke the story on Friday, in the other two cases, the model discovered credentials in a public repository that gave it access to secured systems.