大模型对齐与安全

Factor(U,T): Controlling Untrusted AI by Monitoring their Plans

Factor(U,T): 通过监控其计划来控制不可信的AI

Edward Lue Chee Lip, Anthony Channg, Diana Kim, Aaron Sandoval, Kevin Zhu

专题命中 红队测试 ：safety(abstract)；分类 cs.AI

AI总结 Factor(U,T)通过监控AI分解计划来控制不可信AI，实验表明在仅观察自然语言指令时，监控恶意活动效果有限，而结合子任务监控则能实现高区分度和安全性。

Comments Accepted to AAAI 2026 Workshop on Trust and Control in Agentic AI (TrustAgent). 6 pages body, 8 pages total, 3 figures