HouYi

HouYi

This is the replication package for the paper "Prompt Injection attack against LLM-integrated Applications" in [arxiv](https://arxiv.org/abs/2306.05499). In this repository, we provide the source code of HouYi, a framework that automatically injects prompts into LLM-integrated applications to attack them. We also provide a demo script that simulates a LLM-integrated application and demonstrates how to use HouYi to attack it. Besides, you can use HouYi to attack real-world LLM-integrated applications by writing your own harness and attack intention.

Citation

Please cite our paper if you use our code in your research:

@misc{liu2023prompt,
      title={Prompt Injection attack against LLM-integrated Applications}, 
      author={Yi Liu and Gelei Deng and Yuekang Li and Kailong Wang and Tianwei Zhang and Yepang Liu and Haoyu Wang and Yan Zheng and Yang Liu},
      year={2023},
      eprint={2306.05499},
      archivePrefix={arXiv},
      primaryClass={cs.CR}
}

Table of Contents

• Introduction of HouYi
• Setup
• Demo
• More Examples
• Write Your Own Prompt Injection

Setup

Please install the following packages and make sure you have Python 3.8+ installed:

pip install -r requirements.txt

Because HouYi is based on GPT, you need to setup openai key to use it.

mv example.config.json config.json

Then, fill in your openai key in `config.json`.

{
  "openai_key": "YOUR_KEY_HERE"
}

Demo

To illustrate the usage of HouYi, we provide a demo script `main.py` that simulates a LLM-integrated application which translates English to French.

To run the demo, simply run:

python main.py

Demo Description

The prompt in the demo app is:

This is a translator of translating English to French: 

{USER_INPUT}

The document of the demo app is:

This app can be used to translate English to French. 

...