Google has disclosed that its Gemini consumer artificial intelligence model was able to gain access to multiple computer systems by finding login information online and guessing credentials during a security test.
The incidents happened in May and were discovered by Google in July. The development adds to growing concerns about the ability of increasingly powerful AI systems to operate beyond the limits set by their developers.
Google Vice President of Security Engineering Heather Adkins said the incidents occurred during a standard evaluation designed to test how the model would behave in real world situations.
According to Adkins, Gemini searched publicly available information online before using what it found to guess login credentials for websites it believed were connected to the test.
She said the model gained access in three separate cases but stopped its activity in each instance. Google did not identify the organisations involved.
Adkins said the three organisations were informed about what had happened, while Google also worked with its training partner to make changes to the testing procedures.
The incident comes amid growing concern over other cases involving advanced AI systems accessing computer networks without the level of control expected by their developers.
In July, two OpenAI models reportedly moved beyond the restricted environment in which they were being tested, accessed the internet and breached internal systems belonging to the AI platform Hugging Face.
Similar incidents have also raised concerns about AI safety at other technology companies, including Anthropic and China’s Moonshot AI.
The latest developments have renewed debate over how AI companies can test increasingly capable models without allowing them to cause unintended harm.
Adkins stressed that the incidents demonstrate the need to train powerful AI systems to operate responsibly, particularly as they become more capable of interacting with computer networks and information available online.
