GDM AI Control Roadmap：為假設性的 AI 威脅建立防禦藍圖
當 AI Agent 逐漸具備自主能力，我們該如何防範尚未發生的風險？DeepMind 最新研究提出 TRAIT&R 分類法，搭配 D1–D4 偵測與 R1–R3 回應的分層防禦機制，為企業提供一套在威脅真正成形之前就能先行部署、並逐層驗證成效的安全藍圖。
AI Agents
GDM AI Control Roadmap: A Defensive Blueprint for Hypothetical AI Threats
As AI agents gain autonomy, how do you defend against risks that have not happened yet? DeepMind’s TRAIT&R taxonomy maps D1–D4 detection onto R1–R3 response.