Control over agents is a system, not just a prompt

Google DeepMind published an AI Control Roadmap for its internal systems. The approach combines conventional security measures with monitoring of AI actions, treating system-level controls as a complement to training models to behave appropriately.

Context

Intent and authority should be considered separately. A prompt can explain the task while technical permissions define which actions are possible. Logs and clear review points help investigate mistakes that no instruction could enumerate in advance. Reliable control depends on the surrounding system as well as the agent’s response.