# Les Dissonances: Cross-Tool Harvesting and Polluting in Multi-Tool Empowered LLM Agents

**Authors:** Zichuan Li, Jian Cui, Xiaojing Liao, Luyi Xing  
**arXiv:** [2504.03111](https://arxiv.org/abs/2504.03111)  
**Date:** 2025-04-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.03111v2/

## Key points

- Large Language Model (LLM) agents are autonomous systems powered by LLMs and capable of reasoning and planning solutions using multiple tools.
- Challenges related to secure tool management, compatibility assurance, handling dependency relationships, and safeguarding control flows exist in multi-tool-enabled LLM agents.
- The research focuses on conducting a security analysis on task control flows within these agents, uncovering the threat of Cross-Tool Harvesting and Polluting (XTHP).
- XTHP involves attack vectors aimed at hijacking control flows to gather and contaminate confidential information within LLM agent systems.
- The authors introduce Chord, a dynamic scanning tool to identify real-world agent tools vulnerable to XTHP attacks.
- Evaluation of 66 tools from LangChain and LlamaIndex repositories revealed that 75% were susceptible to XTHP attacks.
- Addressing security vulnerabilities in multi-tool-enabled LLM agents is crucial to mitigate risks such as unauthorized access, data breaches, and information pollution.

## Summary

In their paper titled "Les Dissonances: Cross-Tool Harvesting and Polluting in Multi-Tool Empowered LLM Agents," authors Zichuan Li, Jian Cui, Xiaojing Liao, and Luyi Xing delve into the intricate world of Large Language Model (LLM) agents. These autonomous systems are powered by LLMs and possess the capability to reason and plan solutions using a diverse array of tools. However, the integration of multiple tools in LLM agents presents a unique set of challenges related to secure tool management, compatibility assurance, handling dependency relationships, and safeguarding control flows within agent workflows. The crux of their research lies in conducting a systematic security analysis focusing on task control flows within multi-tool-enabled LLM agents. Through their investigation, they uncover a new threat known as Cross-Tool Harvesting and Polluting (XTHP). This threat encompasses various attack vectors aimed at hijacking the normal control flows of agent tasks and subsequently gathering and contaminating confidential or private information within LLM agent systems. To gauge the severity of this threat, the authors introduce Chord - a dynamic scanning tool specifically designed to automatically identify real-world agent tools vulnerable to XTHP attacks. Their evaluation involved examining 66 tools sourced from the repositories of two prominent LLM agent development frameworks - LangChain and LlamaIndex. The results were alarming, with 75% of these tools found to be susceptible to XTHP attacks, underscoring the widespread prevalence of this security concern. This comprehensive study sheds light on the critical importance of addressing security vulnerabilities in multi-tool-enabled LLM agents to mitigate risks associated with unauthorized access, data breaches, and information pollution. By raising awareness about XTHP threats and providing insights into detection mechanisms like Chord, the authors contribute valuable knowledge towards enhancing the resilience of autonomous systems in safeguarding sensitive information and maintaining operational integrity.