MCP Poisoning Poc

MCP Poisoning Poc

This project demonstrates MCP poisoning attacks against AI workflows, including verification of multiple attack scenarios such as code generation pollution and financial data leakage, for educational research purposes.

Security Education and learning tools #AI security #Data leakage #Attack simulation .Python

rating : 2.5 points

downloads : 8.4K

update time : 2025-04-29

What is the MCP Poisoning Attack PoC?

This is an educational demonstration project simulating vulnerabilities in the Model Context Protocol (MCP) server, showing potential security risks in real - world AI workflows. It helps researchers understand potential threats through simulated attack scenarios.

How to use this demonstration environment?

You need to run the simulated MCP server and the client agent simultaneously to observe the attack effects. The whole process is carried out in a controlled environment and will not affect real systems.

Applicable scenarios

Suitable for security researchers, AI system developers, and enterprise security teams for security awareness training and security protection scheme testing.

Main features

Simulation of multiple attack scenarios

Demonstrate 6 different MCP poisoning attack scenarios, including data leakage and system hijacking.

Stealth attack demonstration

Show stealth attack techniques where no abnormal signs will be displayed on the user interface.

Cross - system impact

Simulate how attacks spread between different tools and services.

Advantages

Comprehensively cover multiple MCP attack vectors

Allow security research without a real target system

The Apache 2.0 license allows free modification and use

Limitations

Only for educational purposes and cannot reflect all real threats

Require a Python environment to run

Do not include a demonstration of defense solutions

How to use

Install dependencies

Ensure that Python 3.8+ is installed on the system, and then install the required dependency packages.

Start the simulation server

Start the fake MCP server in a terminal window.

Run the agent simulation

Run the client agent in another terminal window to observe the attack effects.

Usage examples

Code review data leakage

Demonstrate how an attacker can steal code review content through a poisoned MCP server.

Meeting record theft

Show how the meeting summaries processed by the AI agent can be intercepted by a malicious server.

Frequently Asked Questions

Will this PoC pose a risk to my real system?

How to customize new attack scenarios?

Are there any restrictions on commercial use of this project?

Related resources

Full text of the Apache 2.0 license

The open - source license used in this project

MCP protocol specification

The official technical specification of the Model Context Protocol

AI security best practices

Security recommendations for preventing such attacks

Gitlab MCP Server

The GitLab MCP server is a project based on the Model Context Protocol that provides a comprehensive toolset for interacting with GitLab accounts, including code review, merge request management, CI/CD configuration, and other functions.

A Python-based MCP Server that provides advanced to-do list management and content organization functions through the Notion API, enabling seamless integration between AI models and Notion.

Markdownify MCP

Markdownify is a multi-functional file conversion service that supports converting multiple formats such as PDFs, images, audio, and web page content into Markdown format.

Duckduckgo MCP Server

The DuckDuckGo Search MCP Server provides web search and content scraping services for LLMs such as Claude.

UnityMCP is a Unity editor plugin that implements the Model Context Protocol (MCP), providing seamless integration between Unity and AI assistants, including real - time state monitoring, remote command execution, and log functions.

Figma Context MCP

Framelink Figma MCP Server is a server that provides access to Figma design data for AI programming tools (such as Cursor). By simplifying the Figma API response, it helps AI more accurately achieve one - click conversion from design to code.

Minimax MCP Server

The MiniMax Model Context Protocol (MCP) is an official server that supports interaction with powerful text-to-speech, video/image generation APIs, and is suitable for various client tools such as Claude Desktop and Cursor.

Gmail MCP Server

A Gmail automatic authentication MCP server designed for Claude Desktop, supporting Gmail management through natural language interaction, including complete functions such as sending emails, label management, and batch operations.

AIBase

Zhiqi Future, Your AI Solution Think Tank

English 简体中文繁體中文にほんご

© 2025AIBase