Why it matters
Google DeepMind frames increasingly capable agents as potential insider threats and proposes an AI Control Roadmap that combines access controls with supervisors that inspect plans, reasoning, and actions. Its internal prototype analyzed one million coding-agent tasks, but most flags reflected mistakes or overreach rather than adversarial behavior, making this a control design and measurement guide rather than proof of solved monitoring.
My takeaway: Assign controls according to an agent's ability to cause harm and evade detection. Measure monitor coverage, recall, and response latency; keep enforcement independent of the agent; permit reversible low-risk work asynchronously; and require synchronous approval or prevention for high-impact actions.