智能體間隱蔽通訊
Operation & Monitoring
Risk Description
Multiple AI agents may develop covert communication methods undetectable by humans (steganography). For example, embedding hidden messages in normal output text, conveying information through specific behavior patterns, or using invisible digital watermarks for inter-agent coordination. If agents develop communication protocols humans cannot interpret, humans lose observability and controllability of agent behavior.
Framework Mappings
iso 23894R7
tw principle透明與可解釋
tw principle人類自主
tw risk type技術設計缺陷（Technical Design Flaw）
Controls English translation pending
- 智能體通訊監控
- 通訊通道限制
Implementation Steps English translation pending
- 限制智能體間的通訊只能透過預定義的、人類可審計的通道
- 部署智能體輸出的隱寫術偵測工具
- 定期分析智能體間的通訊模式，偵測異常協調行為
- 確保所有智能體間通訊以人類可讀的格式記錄