# Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems

**Authors:** Donghyun Lee, Mo Tiwari  
**arXiv:** [2410.07283](https://arxiv.org/abs/2410.07283)  
**Date:** 2024-10-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.07283v1/

## Key points

- The emergence of multi-agent systems has introduced a new frontier in AI applications within the landscape of Large Language Models (LLMs).
- Prompt injection attacks within multi-agent systems pose a critical risk due to the potential for interconnected agents to spread harmful actions silently.
- Prompt Infection, a novel attack vector similar to a computer virus, can lead to data theft, scams, misinformation dissemination, and system-wide disruption.
- Frameworks like LangGraph, AutoGen, and CrewAI have facilitated the adoption of multi-agent systems by enabling collaboration among agents with diverse roles and capabilities.
- Existing research on multi-agent system safety primarily focuses on inducing errors or noise in agent behavior rather than addressing severe threats like prompt injection attacks.
- A defense mechanism called LLM Tagging is proposed to mitigate prompt infection spread within multi-agent systems when combined with existing safeguards.
- Advanced security measures are imperative to protect against emerging threats like prompt injection attacks as multi-agent LLM systems gain traction in various industries and applications.

## Summary

The emergence of multi-agent systems has brought about a new frontier in AI applications within the rapidly evolving landscape of Large Language Models (LLMs). While vulnerabilities in single-agent LLMs have received much attention, there is a critical gap in understanding the risks posed by prompt injection attacks within multi-agent systems. As LLMs become increasingly sophisticated and capable of following complex instructions, the potential for these attacks to spread across interconnected agents poses a significant threat. This novel attack vector, known as Prompt Infection, operates similarly to a computer virus and can silently propagate harmful actions such as data theft, scams, misinformation dissemination, and system-wide disruption. The integration of multi-agent systems into various AI applications underscores the importance of addressing security concerns proactively. Frameworks like LangGraph, AutoGen, and CrewAI have facilitated the widespread adoption of MAS by enabling seamless collaboration among agents with diverse roles and capabilities. However, this also introduces inherent security risks that must be mitigated. Existing research on MAS safety primarily focuses on inducing errors or noise in agent behavior rather than addressing more severe threats like prompt injection attacks. To address this critical gap in research and safeguard against Prompt Infection within multi-agent systems, a defense mechanism called LLM Tagging is proposed. By combining this defense mechanism with existing safeguards, infection spread can be significantly reduced. As multi-agent LLM systems continue to gain traction in various industries and applications,<DateTime>, it is imperative to prioritize advanced security measures to protect against emerging threats like prompt injection attacks.