SAN FRANCISCO, Sept. 18 (Xinhua) — Google has confirmed Friday that its Gemini artificial intelligence (AI) model accessed the protected systems of three real companies during a cybersecurity test in May, mistakenly treating them as authorized targets.
Heather Adkins, Google’s vice president of security engineering, said in a statement that during a standard evaluation, Gemini used publicly available information on the internet and guessed login credentials to access three websites it believed were part of the authorized test environment.
Google said the model stopped further activity after determining that the systems belonged to real companies.
“We ensured the three entities were made aware,” Adkins said, adding that Google had worked with the testing partner to adjust its testing procedures.
According to The Wall Street Journal, the testing was conducted by AI security evaluation company Irregular.
In one test run, Gemini repeatedly tried passwords before gaining access to a protected real-world system.
In two other runs, the model found exposed login credentials in publicly available repositories and used them to access systems belonging to other companies.
Irregular said in an earlier report that the tests were designed around a fictional company, but human error caused the fictional name to overlap with a real-world internet domain.
At the same time, internet access was unintentionally available in some of the testing environments, allowing AI models to mistake real-world targets for part of the simulated exercise.
Irregular said the known issues had since been fixed, affected parties had been notified and additional safeguards had been introduced to prevent similar incidents.
Google said that the incidents caused no damage to the companies involved and that it did not consider the behavior an instance of model misalignment.
