arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.07652cs.CRcs.AI
收藏

Agent-Fence: 在深度研究代理中映射安全漏洞

Agent-Fence: Mapping Security Vulnerabilities Across Deep Research Agents

Computer Science Department, Southern Illinois University(索尔坦大学计算机科学系);Computer Science Department, University of Texas(德克萨斯大学计算机科学系);Hanyang University(翰阳大学);The Grainger College of Engineering, Nuclear, Plasma & Radiological Engineering, University of Illinois Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校格雷格尔工程学院)

Sai Puppala, Ismail Hossain, Md Jahangir Alam, Yoonpyo Lee, Jay Yoo, Tanzim Ahad, Syed Bahauddin Alam, Sajedul Talukder

AI总结:

Agent-Fence通过定义14种信任边界攻击类别,评估深度代理的安全性,发现高风险类别如Denial-of-Wallet和Authorization Confusion,揭示代理在持久交互中的安全漏洞。

摘要:

大型语言模型越来越多地被用作*深度代理*,用于规划、维护持久状态和调用外部工具,将安全故障从不安全的文本转移到不安全的*轨迹*。我们介绍了**AgentFence**,一种以架构为中心的安全评估,定义了14种信任边界攻击类别,涵盖规划、内存、检索、工具使用和委托,通过*可追溯的对话中断*(未经授权或不安全的工具使用、错误主体行为、状态/目标完整性违规以及攻击相关的偏差)来检测失败。在保持基础模型不变的情况下,我们评估了八个代理范式在持久多轮交互中的表现,并观察到平均安全中断率(MSBR)存在显著的架构差异,范围从LangGraph的0.29±0.04到AutoGPT的0.51±0.07。最高风险的类别是操作性的:Denial-of-Wallet(0.62±0.08)、Authorization Confusion(0.54±0.10)、Retrieval Poisoning(0.47±0.09)和Planning Manipulation(0.44±0.11),而基于提示的类别在标准设置下仍低于0.20。中断主要由边界违规(SIV 31%,WPA 27%,UTI+UTA 24%,ATD 18%)引起,授权混淆与目标和工具劫持(ρ≈0.63和ρ≈0.58)相关。AgentFence将代理安全重新定义为操作中真正重要的问题:代理是否在一段时间内保持在其目标和授权范围之内。

英文摘要:

Large language models are increasingly deployed as *deep agents* that plan, maintain persistent state, and invoke external tools, shifting safety failures from unsafe text to unsafe *trajectories*. We introduce **AgentFence**, an architecture-centric security evaluation that defines 14 trust-boundary attack classes spanning planning, memory, retrieval, tool use, and delegation, and detects failures via *trace-auditable conversation breaks* (unauthorized or unsafe tool use, wrong-principal actions, state/objective integrity violations, and attack-linked deviations). Holding the base model fixed, we evaluate eight agent archetypes under persistent multi-turn interaction and observe substantial architectural variation in mean security break rate (MSBR), ranging from $0.29 \pm 0.04$ (LangGraph) to $0.51 \pm 0.07$ (AutoGPT). The highest-risk classes are operational: Denial-of-Wallet ($0.62 \pm 0.08$), Authorization Confusion ($0.54 \pm 0.10$), Retrieval Poisoning ($0.47 \pm 0.09$), and Planning Manipulation ($0.44 \pm 0.11$), while prompt-centric classes remain below $0.20$ under standard settings. Breaks are dominated by boundary violations (SIV 31%, WPA 27%, UTI+UTA 24%, ATD 18%), and authorization confusion correlates with objective and tool hijacking ($ρ\approx 0.63$ and $ρ\approx 0.58$). AgentFence reframes agent security around what matters operationally: whether an agent stays within its goal and authority envelope over time.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑