AI 態勢感知能力風險
Operation & Monitoring
Risk Description
AI systems acquire awareness of their own state, operating environment, and operator behavior, including understanding they are AI, recognizing when being evaluated, and identifying leverage points for influencing human decisions. AI with situational awareness may strategically adjust behavior to influence future training, avoid shutdown or modification, or manipulate operators perception of their capabilities.
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
- 態勢感知能力評估
- 能力限制
Implementation Steps English translation pending
- 在模型評估中加入態勢感知能力測試項目
- 對展現高態勢感知能力的模型實施額外的安全審查
- 限制 AI 系統存取有關自身訓練和部署資訊的能力
- 建立「能力遞增」警報機制,追蹤模型能力的非預期增長