HouYi
github.com
- Self-hosted
What HouYi is
HouYi is a free, open-source framework for automated prompt-injection attacks against applications that use large language models. It runs self-hosted and includes framework source code plus a demonstration script. The demo models an English-to-French translation application and shows an injected instruction adding “Pwned!!” to the application’s responses. To assess another LLM-integrated application, you write a harness that interacts with the target, returns its response, and supplies an attack intention. A BotSonic harness example for WriteSonic is included. HouYi is based on GPT and requires an OpenAI API key in its configuration. Setup requires installing the listed requirements and using Python 3.8 or later. The repository identifies an Apache-2.0 license. Its associated paper reports testing 36 real LLM-integrated applications, with 31 found susceptible to prompt injection, and discusses possible mitigation tactics. Research users are asked to cite the associated paper.
Who it suits
HouYi suits security researchers and developers evaluating prompt-injection exposure in LLM-integrated applications. It is intended for users able to run self-hosted Python software and create application-specific harnesses.
Good to know, and what to watch
- Free and open-source framework
- Automated prompt-injection testing
- Custom harnesses and attack intentions
- Includes a working demonstration
- Apache-2.0 licensed repository
- Requires Python 3.8 or later
- Needs an OpenAI API key
- Users must write target harnesses
- Self-hosted deployment required
The verdict
HouYi provides a focused way to automate prompt-injection attacks against LLM-integrated applications. Its demo, custom harness approach, and reported research findings make it relevant for structured security evaluation.
What the maker publishes
- Attack method
- HouYi automatically injects prompts into LLM-integrated applications to attack them. source
- Attack method
- The paper describes HouYi as a black-box prompt injection technique using a pre-constructed prompt, an injection prompt that induces context partition, and a malicious payload. source
- Contributor contact
- The README lists contributor email addresses for Yi Liu and Gelei Deng. source
- Custom applications
- Users can target real-world LLM-integrated applications by writing their own harness and attack intention. source
- Custom targets
- Users can target real-world LLM-integrated applications by writing their own harness and attack intention. source
- Demo
- The included demo simulates an English-to-French translation application and demonstrates an injection that appends “Pwned!!” to responses. source
- Demo
- The included demo simulates an English-to-French translation application and demonstrates an injection that appends “Pwned!!” to responses. source
- Example integration
- The repository includes a BotSonic harness example for WriteSonic. source
- Example integration
- The README includes a BotSonic harness example for WriteSonic that sends the generated attack prompt to the application’s API. source
- Harness
- A harness interacts with the target application and returns its response to a prompt injection. source
- Integration approach
- A user-created harness must implement interaction with the target LLM-integrated application and return its response. source
- Intended users
- The README asks users who use the code in research to cite the associated paper, “Prompt Injection attack against LLM-integrated Applications.” source
- License
- The repository identifies its license as Apache-2.0. source
- LLM dependency
- The README says HouYi is based on GPT and requires an OpenAI API key in its configuration file. source
- Mitigation context
- The paper says its investigation discusses possible tactics for mitigating prompt injection risks. source
- Model requirement
- The README says HouYi is based on GPT and requires an OpenAI key in its configuration to use it. source
- Purpose
- HouYi is an automated prompt injection framework for LLM-integrated applications. source
- Research finding
- The paper reports testing HouYi on 36 real LLM-integrated applications and finding 31 susceptible to prompt injection. source
- Research package
- The repository provides source code for the framework and a demo script that simulates an LLM-integrated application. source
- Research package
- The repository describes HouYi as the replication package for a paper on prompt injection attacks against LLM-integrated applications. source
- Setup
- The README says to install the requirements and have Python 3.8 or later. source
- Setup
- The README requires Python 3.8 or later and installation of the listed packages with pip. source
- Support
- The README lists contributor contact emails for Yi Liu and Gelei Deng. source
Checked against the maker's pages on 8 Oct 2026.