AI Safety and Control

如何讓越來越自主的 AI 代理維持對齊、受界限約束、且可被糾正——涵蓋對齊與控制機制、沙盒隔離與權限設計、多代理群體湧現的系統性風險、架構級失敗模式，以及自主代理打開的資安面（對抗性攻擊、零日漏洞的自動化利用）。核心問題：當代理接手越來越多「怎麼做」，人類要如何保住「定義界限、驗證行為、在系統偏航時把它拉回來」的能力。

Keeping increasingly autonomous AI agents aligned, bounded, and correctable — spanning alignment/control, sandboxing and permissioning, emergent multi-agent risk, architecture-level failure modes, and the security surface opened by autonomous agents. The core question: as agents do more of the “how,” how do humans retain the power to set limits, verify behaviour, and pull the system back when it drifts.