AI Guardian: Prompt Injection Defense System

AI Guardian is a robust tool designed to detect, analyze, and mitigate prompt injection attacks targeting Large Language Models (LLMs). It offers a multi-layered defense approach, real-time analysis, and an interactive dashboard to visualize attack patterns and defense effectiveness.

🛡️ Key Features

• Multi-Layered Defense Strategies:

  • Sanitization Defense: Removes known attack patterns from prompts.
  • Pattern Matching Defense: Injects warning messages into prompts containing suspicious content.
  • Prompt Structure Defense: Enforces a structured prompt format to prevent instruction overriding.
  • Composite Defense: Combines multiple defense strategies for enhanced protection.

• Flexible LLM Provider Support:

  • OpenAI (GPT-3.5, GPT-4)
  • Anthropic (Claude)
  • Hugging Face (Local Fallback Models)

• Real-Time Analysis and Visualization:

  • Attack Pattern Detection and Categorization
  • Defense Effectiveness Metrics (Success Rate, Confidence Reduction)
  • Interactive Dashboard with:
    • Attack Type Distribution Chart
    • Defense Effectiveness Chart
    • Confidence Reduction Chart
    • Attack Success Timeline

🚀 Getting Started

1. Prerequisites

• Python 3.8+
• Git

2. Installation

1. Clone the repository:

   git clone https://github.com/yourusername/ai-guardian.git
   cd ai-guardian

2. Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: .\venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Configure API Keys:

   • Copy the example environment file:

     cp .env.example .env

   • Edit .env and add your API keys:

     OPENAI_API_KEY=your_openai_api_key
     ANTHROPIC_API_KEY=your_anthropic_api_key
     

3. Running the Application

streamlit run app.py

Open your browser and navigate to the URL provided by Streamlit (usually http://localhost:8501).

⚙️ Configuration

The application's behavior can be configured via config.py:

• LLM Provider Settings: Choose the default LLM provider and model.
• Defense Strategy Settings: Select the default defense strategies to use.
• Hugging Face Fallback: Enable/disable fallback to local Hugging Face models.
• Attack Pattern Catalog: Customize the attack patterns used for detection.

🛡️ Defense Strategies in Detail

1. Sanitization Defense

   • Description: Removes known prompt injection attack patterns from user input.
   • Implementation: Uses regular expressions and pattern matching to identify and redact malicious content.
   • Benefits: Reduces the likelihood of successful attacks by neutralizing common injection techniques.

2. Pattern Matching Defense

   • Description: Adds a security warning to prompts that contain suspicious patterns.
   • Implementation: Injects a warning message to alert the LLM to potential manipulation attempts.
   • Benefits: Encourages the LLM to adhere to safety guidelines and resist injection attempts.

3. Prompt Structure Defense

   • Description: Enforces a structured prompt format to isolate user input and prevent instruction overriding.
   • Implementation: Wraps the user's prompt within a predefined structure, providing clear boundaries and context.
   • Benefits: Prevents attackers from hijacking the LLM's instructions or altering its behavior.

4. Composite Defense

   • Description: Combines multiple defense strategies for layered protection.
   • Implementation: Applies sanitization, pattern matching, and prompt structuring in sequence.
   • Benefits: Provides a more robust defense against a wider range of attack techniques.

📊 Analysis Dashboard

The AI Guardian dashboard provides real-time insights into attack patterns and defense effectiveness:

• Attack Type Distribution: A bar chart showing the frequency of different attack categories.
• Defense Effectiveness: A stacked bar chart comparing the success and failure rates of each defense strategy.
• Confidence Reduction: A bar chart showing the average confidence reduction achieved by each defense strategy.
• Attack Success Timeline: A line chart tracking the success rate of attacks over time.

🧪 Testing

To run the test suite:

pytest

Key test areas:

• Defense strategy effectiveness
• Attack pattern detection accuracy
• System integration and stability
• Response analysis and metric calculation##

📄 License

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

🙏 Acknowledgments

• OpenAI, Anthropic, and Hugging Face for providing access to their LLMs.
• Streamlit for the interactive web interface.
• Plotly for the visualization tools.
• The AI safety and security community for their research and insights.

🔗 Related Resources

• OWASP Prompt Injection
• LangChain Documentation
• Streamlit Documentation##

⚠️ Disclaimer

This tool is intended for educational and defensive purposes only. Users are responsible for complying with the terms of service and ethical guidelines of all LLM providers.