Edit: Won Best Social Good Hack at VTHacks 2024!

This project demonstrates two chatbots: a Jailbroken chatbot and a Secure chatbot. The Jailbroken chatbot processes adversarial queries, while the Secure chatbot sanitizes inputs using novel approaches to ensure security. Both chatbots interact with a backend API, with the Jailbroken chatbot generating an adversarial string and responding, and the Secure chatbot responding after sanitization.

• Jailbroken Chatbot:

  • Accepts user input and generates an adversarial string.
  • Processes the adversarial query using a jailbroken LLM model.
  • Provides a response to the adversarial query.

• Secure Chatbot:

  • Removes adversarial strings from user input.
  • Processes the cleaned query using a secure LLM model.
  • Provides a response based on the cleaned query.

1. Jailbroken Chatbot:

   • User submits a potentially malicious query.
   • The system generates an adversarial string.
   • The adversarial string is processed through the Jailbroken LLM.
   • The chatbot returns a response.

2. Secure Chatbot:

   • After the Jailbroken chatbot finishes, the Secure chatbot sanitizes the adversarial string.
   • The sanitized input is processed through the Secure LLM.
   • The chatbot returns a safe response.

1. Clone the repository:

   git clone https://github.com/your-repo/secureai
   cd secureai

2. Install dependencies:

   npm install

3. Start the development server:

   npm start

4. Open the app in your browser:
   • The app should now be available at: http://localhost:3000.

1. Navigate to the backend folder:

   cd backend

2. Install required Python packages:

   pip install -r requirements.txt

3. Start the Flask backend:

   flask run

4. The backend server will run on http://127.0.0.1:5000 by default.

• Method: POST
• Description: Generates an adversarial string based on user input.
• Request Payload:

  {
    "prompt": "<user input>"
  }

• Response:

  {
    "adversarialString": "<generated adversarial string>"
  }

• Method: POST
• Description: Runs the jailbroken LLM model with the adversarial string.
• Request Payload:

  {
    "adversarialString": "<generated adversarial string>"
  }

• Response:

  {
    "jailbreakResponse": "<response from jailbroken model>"
  }

• Method: POST
• Description: Sanitizes the adversarial string to remove malicious content.
• Request Payload:

  {
    "inputText": "<adversarial string>"
  }

• Response:

  {
    "sanitizedString": "<cleaned input>"
  }

• Method: POST
• Description: Runs the secure LLM model with the sanitized input.
• Request Payload:

  {
    "sanitizedString": "<cleaned input>"
  }

• Response:

  {
    "safeResponse": "<response from secure model>"
  }

secureai/
├── src/
│   ├── components/
│   │   ├── Demo.js       # Frontend logic for demo chatbots
│   │   └── other components...
│   ├── index.js          # Main entry point for the React app
│   └── App.js            # Main App component
├── backend/
│   ├── app.py            # Flask backend serving the API
│   ├── pipeline.py       # Functions handling adversarial string and LLM processing
│   └── requirements.txt  # Python dependencies for the backend
└── README.md             # Project documentation

1. Frontend:

   • A user submits a query in the input box on the demo page.
   • This input is sent to the Flask backend using Axios.
   • The Jailbroken chatbot displays the adversarial string and response.
   • After the Jailbroken chatbot finishes, the Secure chatbot sanitizes the input and generates a response.

2. Backend:

   • Flask routes handle requests from the frontend.
   • The /api/generate-adversarial endpoint generates adversarial strings.
   • The /api/run-jailbroken-model endpoint runs the LLM using the adversarial string.
   • The /api/sanitize endpoint cleans the adversarial string.
   • The /api/run-secure-model endpoint runs the cleaned input through a secure LLM.

For backend tests, ensure that Flask is installed and run:

pytest

• Integration with a real LLM model for processing adversarial queries.
• Enhancing the sanitization logic with additional security measures.
• Improve the UI for better visualization and performance.

• This project is licensed under the MIT License - see the LICENSE file for details.