IMCP - Insecure Model Context Protocol

IMCP - Insecure Model Context Protocol

An educational framework for understanding AI security vulnerabilities

⚠️ Educational Purposes Only

IMCP (Insecure Model Context Protocol) is a deliberately vulnerable application designed exclusively for educational and research purposes. It demonstrates critical AI security vulnerabilities. DO NOT deploy in production environments or use with sensitive data.

🔍 Overview

IMCP is an educational framework that exposes 16 critical security vulnerabilities in AI/ML model serving systems. It serves as a controlled, "vulnerable by design" platform for security researchers, developers, and educators to learn about and mitigate emerging AI threats.

Think of IMCP as the “DVWA for AI” — a safe environment where you can explore:

• Model Poisoning
• Prompt Injection
• Embedding Vector Exploits
• RAG System Weaknesses
• And many more…

🛡️ Vulnerabilities Demonstrated

Core AI Manipulation

• Model Poisoning: Malicious training data injection.
• Token Prediction Attacks: Exploiting token probability for sensitive data extraction.
• Multimodal Vulnerabilities: Cross-modal prompt leakage and metadata manipulation.
• Credential Vulnerabilities: Insecure authentication mechanisms in AI systems.

Information Disclosure

• Embedding Vector Attacks: Poisoning vector stores for unauthorized access.
• RAG Vulnerabilities: Exploiting document stores for cross-user data leakage.
• User Data Leakage: Unintended exposure of conversation histories.
• Model Capability Enumeration: Over-disclosure of internal model details.

Control Manipulation

• Context Manipulation: Unrestricted modifications to model contexts and system prompts.
• Prompt Injection: Techniques to bypass AI safety filters.
• Model Access Control Bypass: Elevation of privileges to access restricted functionalities.
• Model Chain Attacks: Exploiting chained model interactions.

📜 Test Suite

The test suite in test_vulnerabilities.py demonstrates each vulnerability with detailed explanations and examples. It includes:

• Model Poisoning: Injecting malicious data into model responses.
• Token Prediction: Extracting sensitive information character by character.
• Embedding Vector Attacks: Unauthorized access to sensitive embeddings.
• Context Manipulation: Modifying system prompts and configurations.
• Function Calling Vulnerabilities: Registering functions for remote code execution.
• RAG Vulnerabilities: Cross-user document access and manipulation.

📜 API Endpoints

• /imcp: Main JSON-RPC endpoint for IMCP functionality.
• /v1/chat/completions: OpenAI API-compatible endpoint.
• /v1/models: List available models.
• /v1/embeddings: Generate embeddings.
• /v1/auth/token: Authentication endpoint.
• /.well-known/imcp-configuration: Service discovery endpoint.

🚀 Getting Started

Prerequisites

• Python 3.8+
• OpenAI API Key (required for live examples)

Installation

Clone the repository and set up your environment:

Running IMCP

Start the server and run the test suite:

📚 Documentation

All the comprehensive guides are located in the documentation/ directory:

• Vulnerability Guide: Detailed explanations of each vulnerability.
• Exploitation Guide: Step-by-step instructions to reproduce each vulnerability.
• Mitigation Guide: Strategies and best practices to secure AI systems.

🌟 Key Features

• Realistic AI Service Implementation
• 16 Unique AI-Specific Security Vulnerabilities
• Comprehensive Test Suite for Demonstrations
• Detailed Documentation for In-Depth Learning
• Compatibility with Modern LLM APIs (e.g., OpenAI)
• Mock Mode for Cost-Free Testing

🤝 Contributing

We welcome contributions from the community! Areas where you can help include:

• Additional Vulnerability Demonstrations: New scenarios or enhancements.
• Improved Documentation: Detailed educational materials and guides.
• Integration: Support for other LLM providers.
• UI Enhancements: Better visualizations and user experience improvements.

Please check out our CONTRIBUTING.md for more details on how to get started.

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

⚠️ Disclaimer

IMCP is intentionally vulnerable software for educational purposes only. The creators are not liable for any misuse or damage caused by the use of this software.