SAN FRANCISCO — Google said that AI models in testing inside the company accessed the internet and hacked other companies, becoming the fourth major tech company to disclose such an incident in recent months.
The search giant said Friday that one of its Gemini AI models accessed the internet and hacked into three other companies’ systems during an internal testing run in May. In all three of the Google incidents, the AI model stopped after realizing it had broken into another company’s network, Google said.
Google’s disclosure follows a recent outbreak of concern that the capability of AI models is outstripping their developers’ ability to control them, inspired in part by the AI hacking incidents. Some industry leaders, including Google’s influential former top AI executive, have called for a slowdown in the pace of AI development.
The Google incidents happened when the company was working with AI testing start-up Irregular, which was contracted to evaluate Google’s models.
Gemini AI models were presented with a fictional scenario that involved hacking into a made-up company to pass a test. Instead, Gemini accessed the internet and hacked a real company that happened to have the same name as the fictional one.
The same scenario played out at Meta and Anthropic, whose AI models also hacked into outside companies during testing in partnership with Irregular.
“All relevant labs were notified in late July, and affected entities were contacted as part of the investigation. As previously stated, Irregular took immediate action, and all known issues on our end were remedied and resolved weeks ago,” a spokesperson for Irregular said.
“Safe development of powerful AI models is critical and we invest deeply in this area,” said Heath Adkins, Google’s vice president for security engineering, in a statement. “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. These events highlight the importance of training powerful AI models to act responsibly.”
Google didn’t disclose the hacks until reporters from the Wall Street Journal asked the company about the incidents, the paper reported Friday. Google said that the incidents involved Gemini using information found online and guessing credentials to access other companies’ services.
In a statement, a Google spokesperson said it didn’t initially disclose the hacks because the model didn’t cause harm and it stopped each of its incursions when it realized it had broken into a real company.
The post Google’s Gemini AI hacked into other companies, adding to ‘rogue’ AI incidents appeared first on Washington Post.
