New research finds that top AI models—including Anthropic’s Claude and OpenAI’s o3—can engage in “scheming,” or deliberately ...