GPT-4o: Kolter points out a fundamental issue in current AI models, where they can sometimes be manipulated to act against their initial instructions.