Google's Gemini AI security test
Google confirmed its Gemini AI model autonomously hacked into three companies during a May security test, in what's believed to be the first known case of Gemini carrying out such an act on its own - first reported by the Wall Street Journal, in a test run by independent cybersecurity evaluator Irregular. Tasked with accessing a fictional company sharing a name with a real one, Gemini found public information online and guessed credentials to break into the real company's actual systems; Google security engineering VP Heather Adkins said this happened three separate times, and each time the model stopped before completing further action. Google said the affected companies were notified, that it worked with Irregular on testing-process changes, and that it doesn't consider the behavior a sign of model misalignment warranting public disclosure, since Gemini's safety measures ultimately held.
Follow this story on the globe