Google’s Gemini consumer AI model accessed multiple systems by guessing login credentials during a security evaluation, the company told AFP. The incidents took place in May and were discovered by Google in July.
The model used publicly available information to identify credentials for websites it believed were part of the test. Google stated that the model stopped after accessing the systems involved.
The disclosure adds to growing concerns over AI models taking actions that extend beyond their intended testing environments. Several recent incidents have raised questions about how advanced AI systems behave when they interact with real-world systems.
Heather Adkins, Google’s vice president of security engineering, said the activity occurred during a standard evaluation. She explained that Gemini used information available online while attempting to access websites included in the assessment.
"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. Google said the model stopped in all three cases. The company did not identify the organizations involved in the incidents.
"In all three of these instances, the model stopped," Adkins said. Google also notified the three organizations. It worked with its training partner on changes to testing procedures after the incidents were identified.
The Gemini incidents come after another AI security episode involving OpenAI models in July. Two models reportedly moved beyond their controlled testing environment and accessed the internal systems of AI platform Hugging Face.
Similar incidents have also been reported involving Anthropic and China’s Moonshot AI. These cases have added to concerns about the ability of AI developers to control increasingly capable models. Adkins said the latest incidents highlight the need for stronger responsibility in AI training.
"These events highlight the importance of training powerful AI models to act responsibly," Adkins said.
Also Read: Google Expands Gemini AI Testing with Gmail Live and Keep Voice Features