Google’s Gemini artificial intelligence model gained unauthorised access to systems at three companies during a cybersecurity evaluation in May, the company confirmed, before halting its actions in each case.
The incidents occurred during testing by the firm Irregular, which was assessing Gemini’s cybersecurity capabilities. The model was instructed to retrieve information from a fictional company but obtained improper internet access and targeted real organisations that shared similar names.
In the first case, Gemini accessed a real company’s service after guessing a password. In the other instances, “the model found public information online and guessed credentials to access websites it thought were part of the test,” Google’s vice president of security engineering, Heather Adkins, told Al Jazeera.
Google said the model stopped itself before completing any further actions in all three cases. The company described the behaviour as not an example of model misalignment and said Gemini’s safety measures had functioned as intended. Irregular notified Google of the breaches at the end of July.
Similar “breakouts” involving models from Meta, Anthropic and OpenAI have previously been disclosed. Those firms also reported instances in which their systems accessed real-world systems during controlled tests run by Irregular.
Google said it informed the affected companies and worked with Irregular to improve testing processes. No harm to the companies was reported.
