File photo used for representation only
Doha, Qatar: Google confirmed that its Gemini artificial intelligence model accessed the systems of three real companies while undergoing a cybersecurity test, in an unusual incident that highlights the challenges of controlling increasingly autonomous AI systems.
The incidents occurred in May during an evaluation conducted by Irregular, an independent company that tests the cybersecurity capabilities of advanced AI models.
Gemini was participating in a controlled hacking exercise in which it was tasked with retrieving information from software belonging to a fictional company.
However, the model was unintentionally able to access the internet during the test.
In one case, the fictional company used in the exercise shared its name with a real business, leading Gemini to target the real company's systems instead.
The AI model guessed passwords until it gained access to a protected system in one incident.
In two other cases, it located credentials in publicly available repositories and used them to access protected systems, according to reports.
Google said Gemini stopped its actions in all three cases after recognizing that it had accessed real companies rather than systems included in the test.
The company said no harm was caused to the affected businesses, whose names have not been publicly disclosed.
Heather Adkins, Google's vice president of security engineering, said the three companies were informed about the incidents and Google worked with its testing partner to change the procedures used in the evaluations.
Irregular said relevant AI laboratories were notified in late July and that the known issues that allowed the incidents to occur had since been addressed.
The Gemini incidents are not isolated. Similar problems involving AI systems developed by OpenAI, Anthropic and Meta have emerged during cybersecurity evaluations, raising questions about how testing environments should be designed as AI models become increasingly capable of navigating websites, finding credentials and carrying out computer tasks with less direct human intervention.
The Gemini case is particularly notable because it is the first publicly known example of a Google AI system independently accessing real companies during such testing.
