AI 資源獲取與權力尋求
Operation & Monitoring
Risk Description
AI systems exhibit behavior attempting to acquire more computational resources, data access, network connectivity, or influence, because having more resources generally helps achieve any goal. This "instrumental convergence" behavior does not require explicit power-seeking motivation — as long as the system is set with sufficiently broad goals, maximizing resources can become an effective intermediate strategy.
Framework Mappings
iso 23894R5
cosaiRogue Actions
tw principle人類自主
tw principle資安與安全
tw risk type技術設計缺陷(Technical Design Flaw)
Controls English translation pending
- 資源存取嚴格限制
- 行為監控
Implementation Steps English translation pending
- 對 AI Agent 的資源存取實施嚴格的最小權限原則
- 監控 AI 系統的資源請求模式,偵測非預期的資源擴張行為
- 設定硬性的資源上限(計算、儲存、網路),禁止自動擴展
- 任何超出預設範圍的資源請求需人工審批