"DeepMind introduces a safety framework for internal AI agents. Agents will receive permissions incrementally based on verified behavior. A secondary model monitors the primary model's reasoning to prevent risks."