面向安全加固的智能体AI遏制架构
原标题：Agentic AI Containment Architecture for Security Hardening
AI 摘要
本文提出了一种面向多智能体AI系统的安全加固架构，将安全视为架构属性，通过六种约束（职责分离、部署前一致性检查、价值流绑定、时间隔离、知识验证、结构完整性验证）强制执行Propose–Verify–Act–Verify执行模型。该架构将系统分析工件映射为可验证的契约，以防御提示注入、编排器操纵、跨会话状态污染和智能体共谋等威胁。通过简历筛选案例展示了在对抗条件下产生可审计、合规结果的能力，并明确区分结构完整性与语义安全。
正文节选
[Page 1] Agentic AI Containment Architecture for Security Hardening Mohamed ElBendary FasTrak SoftWorks, Mequon, WI, USA me@prosterk.com Abstract. Multi-agent AI systems are increasingly deployed in contexts where autonomous coordination, tool use, and continuous learning introduce novel se- curity and governance risks. The containment approach to multi-agent system securit