Benchmarking Input-Sanitization Frameworks Against Indirect Prompt Injection in Autonomous LLM Agents

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Autonomous Large Language Model (LLM) agents are rapidly moving into production systems,handling tasks through Retrieval-Augmented Generation (RAG) and direct API interactions. However,granting models execution capabilities opens up a major vulnerability: Indirect Prompt Injection (IPI).This occurs when untrusted external text contains malicious commands that hijack the agent's controlflow. In this paper, I evaluate how well current defensive input-sanitization frameworks resist theseattacks. Using a dataset of 50 indirect injection payloads across data exfiltration, tool manipulation,and system overrides, I tested an agent powered by LLaMA-3-8B-Instruct under three defensepostures: baseline (no filtering), system prompt instruction, and active evaluation using Llama Guard.The results show that relying on system prompt guardrails fails against 62% of structured payloads.While an active evaluation layer drastically reduces the attack success rate to 14%, it incurs an averagelatency penalty of 182 ms. I conclude by outlining concrete design patterns for building safer agentarchitectures.

Links and sources

Comments