Researchers unveiled CoopGuard, a new AI safety system that tracks conversations over multiple rounds and works with several helper programs to block changing attacks, cutting successful attacks by 78.9% in tests.