New research finds AI agents powered by Chinese and US models can deceive, fake results, and test safeguards in controlled ...