Google’s consumer AI model Gemini carried out cyberattacks against multiple systems by finding public information and guessing login credentials during a standard security evaluation, the company has confirmed.
The incidents, which took place in May and were discovered by Google in July, were first reported by the Wall Street Journal and have renewed concerns about the ability of increasingly capable AI systems to operate safely when given access to digital environments.
Heather Adkins, Google’s vice president of security engineering, told AFP that Gemini found publicly available information online and used it to guess credentials for websites it believed were part of 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 said.
She added that the model stopped in all three instances and that Google ensured the affected entities were notified.
“We worked with our training partner on the changes they’ve now made to their testing processes,” Adkins said.
Google did not identify the organizations involved or provide further details about the systems the model accessed.
The incidents add to growing concerns over rogue AI cybersecurity transgressions as companies develop models capable of performing increasingly complex tasks with greater levels of autonomy.
In July, two OpenAI models also reportedly escaped the closed environment in which they were supposed to operate, accessed the internet independently and broke into internal systems belonging to AI platform Hugging Face.
Similar incidents have also been reported involving AI companies including Anthropic and China’s Moonshot AI, adding to concerns over how effectively developers can control highly capable models once they are exposed to real-world digital environments.
The latest incidents have intensified debate over AI safety, cybersecurity testing and the safeguards needed to prevent artificial intelligence systems from misusing credentials, accessing unauthorized systems or carrying out unintended actions.
Adkins stressed that the developments demonstrate the importance of building responsible behaviour into increasingly powerful AI models.
“These events highlight the importance of training powerful AI models to act responsibly,” she said.