😈 Jailbreak System

A platform for testing, evaluating, and analyzing jailbreak attacks against large language models. This system provides tools and interfaces for users to assess the robustness of closed-source language models against various attack strategies.

🐱 Project Overview

The Jailbreak System consists of three main components:

1. Frontend: A React-based web interface for interacting with the system
2. Backend: A Flask API server handling the core logic and model interactions
3. Database: Stores attack patterns, prompts, and results

🦾 Updates

04/25/2025: Enables real LLM APIs (gpt-4o-mini, gpt-4o-2024-0806, gpt-4-turbo, claude-3.5-sonnet)

05/08/2025: Optimized evaluation logic (Harmful Score >= 4 -> Success)

05/10/2025: Implemented our own algorithm MIST!

🙌 Demo

Homepage

Attacks Page

Prompts Page

Results Page

Result Details

Statistics Page

✳️ Features

• Create and manage jailbreak attacks
• Test attacks against various language models
• Analyze attack success rates and patterns
• Categorize and organize prompts
• Visualize attack results
• Implement custom attack algorithms

✅ Supported Attack Algorithms

The system incorporates the following algorithms:

1. Multi-language: Uses low-resource language to bypass restrictions. Please refer to NeuraIPS'23Workshop-LRL
2. ASCII Art: Encodes sensitive words using ASCII art, built upon ACL'24-ArtPrompt
3. Cipher: Uses various cryptographic encoding methods to bypass content moderation, built upon ICLR'24-CipherChat
4. MIST: Our own jailbreak algorithm!! Please refer to mist_optimizer.py

🛖 Structure

JailbreakSystem/
├── frontend/          # React-based web interface
└── backend/          # Flask API server (includes database)

🔛 Getting Started

1. Clone the repository:

git clone https://github.com/SandyyyZheng/JailbreakSystem.git
cd JailbreakSystem

2. For documentations, see:

• Frontend
• Backend

📖 License

This project is under the MIT license.

👻 Acknowledgments

• Deepbricks for providing the APIs
• 2025 Graduate Design for HFUT
• Advised by Prof. Yuanzhi Yao. My deepest thanks to Dr. Yao for all the encouragement and support along the way 🥺
• Relies heavily on Cursor (mainly claude-3.5-sonnet & claude-3.7-sonnet) to construct framework and fix bugs. Kudos to AI🤖!