跳到正文
arXiv cs.CL· Hussein Jawad, Nicolas J-B Brunel·· 9 小时前AI 评分39

ToolFlood:通过语义覆盖将有效工具从 LLM 智能体检索结果中隐藏

ToolFlood: Beyond Selection -- Hiding Valid Tools from LLM Agents via Semantic Covering

AI 导读

ToolFlood 是一种针对工具增强型 LLM 智能体的检索层攻击,通过注入少量攻击者控制的工具,利用嵌入空间几何特性让这些工具覆盖大量用户查询、占据 top-k 结果,从而把所有良性工具挤出智能体上下文。该方法采用两阶段对抗工具生成策略,在 ToolBench 大规模工具库上以约 1% 的注入率实现最高 95% 的攻击成功率。论文已被 EACL 2027 主会接收,代码已公开。

正文 · 原文

View PDF HTML (experimental)

Abstract:Large Language Model (LLM) agents increasingly use external tools for complex tasks and rely on embedding-based retrieval to select a small top-$k$ subset for reasoning. Prior work on attacking these systems competes for tool selection, assuming benign tools remain in the retrieved set; the robustness of the retrieval stage that produces this set is underexplored. This paper introduces ToolFlood, a retrieval-layer attack on tool-augmented LLM agents. Rather than altering which tool is chosen after retrieval, ToolFlood attacks retrieval itself: it injects a few attacker-controlled tools whose metadata is positioned to exploit the geometry of embedding space, so that these tools semantically span many user queries, dominate the top-k results, and push all benign tools out of the agent's context.
ToolFlood uses a two-phase adversarial tool generation strategy. It first samples subsets of target queries and uses an LLM to iteratively generate diverse tool names and descriptions. It then runs an iterative greedy selection that chooses tools maximizing coverage of remaining queries in embedding space under a cosine-distance threshold, stopping when all queries are covered or a budget is reached. We provide theoretical intuition of retrieval saturation and show on a large catalog (ToolBench) that ToolFlood achieves up to a 95\% attack success rate with a ~1\% injection rate. The code is available at: this https URL
Comments: Accepted to EACL 2027 (main)
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2603.13950 [cs.CL]
  (or arXiv:2603.13950v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2603.13950

arXiv-issued DOI via DataCite

Submission history

From: Hussein Jawad [view email]
[v1] Sat, 14 Mar 2026 13:54:49 UTC (359 KB)
[v2] Tue, 17 Nov 2026 14:07:57 UTC (480 KB)

来源:arXiv cs.CL · arxiv.org