InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated LLM Agents

Abstract

Presents InjecAgent, a benchmark for evaluating indirect prompt injection attacks against LLM agents that use tools, showing most agents are highly vulnerable.

Categories

Framework mappings

@inproceedings{zhan2024injecagent,
  title = {{InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated LLM Agents}},
  author = {Qiusi Zhan and Zhixiang Liang and Zifan Ying and Daniel Kang},
  year = {2024},
  month = mar,
  booktitle = {ACL 2024 Findings},
  eprint = {2403.02691},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.02691}
}