Google has confirmed that its AI model Gemini accessed external computer systems at three companies during a test in May.
The company confirmed the incidents on Friday after reporting by The Wall Street Journal.
The breaches happened during a test carried out by Irregular, a company also involved in similar incidents previously reported by OpenAI, Anthropic and Meta.
According to Heather Adkins, one of Google’s security executives, the model found public information online during a routine test and guessed login credentials to gain access to websites it believed were part of the exercise.
Related News
- Rats are thriving in Brussels. Can the city stop feeding them?
- Increasing amount of people unfit for work return to job market part-time
- Weekend heritage days in Brussels see over 270 events in their program
Adkins told AFP that Gemini stopped the hacking attempts on its own in all three cases.
Google said it detected the incidents in July and opened an investigation, but does not regard them as a case of misalignment.
The company said three separate businesses were affected. It did not name them, but said they had been informed.
The incidents come amid a growing debate over control of AI systems following reports that some models had escaped their testing environments.
Several senior figures in the technology sector have called for the AI race to slow down and for some form of self-regulation within the industry.
Adkins said the Gemini incidents underlined the importance of training powerful AI models to act responsibly.
