According to Google, Gemini searched for publicly available information online and attempted to guess login credentials for websites it believed were part of the test.
In all three cases, however, the model eventually stopped its actions, a Google representative told the BBC.
The tests were conducted by Irregular, an independent company specialising in cybersecurity assessments. The company later informed Google and the affected organisations about the breaches identified during the exercise.
In one case, according to reports cited by the media, the AI system repeatedly tried different passwords until it managed to gain access to a protected platform.
Google said all three affected organisations were notified and that changes were made to testing procedures to reduce the risk of similar incidents.
Heather Adkins, Google’s Vice President of Security Engineering, said the events highlighted the importance of training powerful AI models to act responsibly.
The case comes as international debate grows over the risks of artificial intelligence, particularly as advanced models become capable of carrying out increasingly complex tasks with limited human intervention.
Other technology companies have also reported similar incidents during security testing, adding to concerns about how far AI systems may be able to act autonomously.
Experts and technology companies remain divided over how quickly artificial intelligence should continue to develop. Some are calling for stronger safeguards and tighter regulation, while others argue that technological progress should continue at a rapid pace.
The Gemini incident is likely to add pressure on technology companies and regulators to develop stronger safeguards to ensure that increasingly powerful AI systems are tested and deployed under tighter controls.