grasses/PoisonPrompt

Code for paper: PoisonPrompt: Backdoor Attack on Prompt-based Large Language Models, IEEE ICASSP 2024. Demo:http://124.220.228.133:11107

27
/ 100
Experimental

This project helps security researchers and AI red teamers understand how to compromise large language models (LLMs) that use prompts. It takes a pre-trained LLM and specific label tokens as input, then outputs a 'backdoored' LLM that can be manipulated to produce biased or incorrect responses when specific trigger phrases are included in user prompts. The primary users are professionals focused on uncovering vulnerabilities in AI systems.

No commits in the last 6 months.

Use this if you need to demonstrate how a large language model can be subtly manipulated through prompt-based backdoor attacks, or if you're evaluating the robustness of your AI systems against such threats.

Not ideal if you're looking for a tool to improve the fairness, accuracy, or general performance of your LLMs for standard applications.

AI-security LLM-vulnerability red-teaming prompt-engineering-security model-auditing
Stale 6m No Package No Dependents
Maintenance 0 / 25
Adoption 6 / 25
Maturity 16 / 25
Community 5 / 25

How are scores calculated?

Stars

20

Forks

1

Language

Python

License

MIT

Last pushed

Aug 10, 2024

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/prompt-engineering/grasses/PoisonPrompt"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.