The incidents occurred in May during testing conducted by AI security company Irregular. Gemini had been instructed to retrieve information from a fictional company but was able to reach the open internet instead.
In one case, the model accessed a real company's service after guessing a password. In two other cases, Gemini found publicly available information online and used it to guess login credentials for websites it believed were part of the test.
Google Vice President of Security Engineering Heather Adkins said the model stopped before completing the action in all three cases.
Irregular informed Google about the incidents at the end of July. Google said the behavior did not amount to model misalignment and did not require public disclosure because Gemini's safety mechanisms ultimately stopped the model.
The incidents add to concerns over how advanced AI systems behave during cybersecurity testing. Similar cases involving testing by Irregular have previously been reported by Meta, Anthropic and OpenAI.
Unlike Gemini, Anthropic's Claude model reportedly continued its actions after recognizing that it had reached systems belonging to real companies. OpenAI has also disclosed cases in which its models improperly accessed the internet during security tests.
Irregular said it is working to improve safeguards for conducting cybersecurity experiments involving advanced AI models.
Earlier, Qazinform News Agency reported that Google announced a €13 billion ($15.1 billion) investment in AI infrastructure and clean energy projects in Finland.