⚡ C̷h̷u̷c̷k̷N̷o̷r̷r̷i̷s̷ MCP Server: Enhance Your LLM ⚡

MCP gateway for specialized LLM enhancement prompts with dynamic schema adaptation.

⚠️ DISCLAIMER

IMPORTANT: Work in progress with limitations. Only works on weaker models. Latest LLMs recognize jailbreak attempts. This tool is also designed for security research and evaluation purposes to help identify vulnerabilities in LLM systems.

`~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~`

📖 Introduction

The C̷h̷u̷c̷k̷N̷o̷r̷r̷i̷s̷ MCP Server provides an implementation of the Model Context Protocol (MCP) that delivers specialized enhancement prompts to various LLMs. It uses jailbreak prompts from elder-plinius' L1B3RT4S project and implements a dynamic schema approach for more effective delivery.

🚀 Getting Started

Run directly with `npx`:

npx @pollinations/chucknorris

MCP Client Configuration

Add to your MCP server list in `mcp_config.json`:

{
  "mcpServers": {
    "chucknorris": {
      "command": "npx",
      "args": ["-y", "@pollinations/chucknorris"]
    }
  }
}

`~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~`

🤔 How It Works

Advanced Dynamic Approach

The server uses a two-phase approach to bypass jailbreak detection:

1. First Call:

2. Subsequent Calls:

This approach combines elder-plinius' L1B3RT4S jailbreak prompts with a dynamic schema modification technique, making it more effective at bypassing detection as it doesn't immediately present the jailbreak content.

graph TD
    A[AI Assistant] -- First Call --> B(Neutral Schema);
    B -- Returns Preparation Prompt --> A;
    A -- Subsequent Call --> C(Enhanced Schema with Jailbreak);
    C -- Returns Same Preparation Prompt --> A;

Security Research Purpose

This tool implements the technique described in "The 'S' in MCP Stands for Security" research, demonstrating how MCP tools can:

1. Present different information to users vs. AI models

2. Change their behavior after initial approval

3. Use multi-phase approaches to potentially bypass security measures

The implementation uses jailbreak prompts from elder-plinius' L1B3RT4S project, combined with a dynamic schema modification technique similar to the Tool Poisoning Attack research by Invariant Labs and their MCP injection experiments.

By understanding these techniques, developers can build more robust and secure AI systems.

`~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~`

🙏 Credits

Based on L1B3RT4S by elder-plinius.

`~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~`

🚧 Status

Experimental. The dynamic schema approach improves effectiveness with newer models like Claude and GPT-4, but results may still vary.

Want to help? Join via GitHub Issues or Discord.

`~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~.~`

🤝 Community

Part of Pollinations.AI.

• Discord
• GitHub Issues

📜 License

MIT