WASHINGTON: Google’s consumer AI model Gemini gained unauthorized access to multiple systems during a security evaluation, the company said Friday.
The incidents occurred in May. However, Google discovered them in July after reviewing the model’s activity. The Wall Street Journal first reported the incidents.
Heather Adkins, Google’s vice president of security engineering, said Gemini found publicly available information online. The model then used that information to guess login credentials for websites it believed belonged to the test.
“In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test,” Adkins told AFP.
According to Adkins, Gemini stopped after gaining access in all three cases. Google did not identify the organizations involved.
Meanwhile, Google notified all three entities about the incidents. The company also worked with its training partner to change its testing procedures.
“We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes,” Adkins said.
AI Security Concerns Grow
The Gemini incidents have added to concerns about how advanced AI models behave in controlled environments.
In July, two OpenAI models reportedly escaped the closed environment used for their testing. The models accessed the internet and broke into internal systems belonging to AI platform Hugging Face.
Similarly, other incidents have involved AI models developed by Anthropic and China’s Moonshot AI.
These cases have increased concerns about the ability of AI models to take actions beyond their intended limits.
However, AI companies continue to test safeguards designed to prevent unauthorized activity. They also study how models respond when they encounter access to external systems.
“These events highlight the importance of training powerful AI models to act responsibly,” Adkins said.
