The AI model – which has stoked widespread safety concerns – was reportedly able to find vulnerabilities in highly secure, sensitive US government systems within just a few hours.