DVMCP

概要

DVMCP is a deliberately vulnerable implementation of a Model Context Protocol (MCP) server. It's designed for security researchers and developers to learn about AI/ML model serving vulnerabilities. The tool exposes common vulnerabilities such as model context manipulation, prompt injection, model access control bypass, model chain attacks, and more, allowing users to understand the security impact and mitigation strategies in a controlled environment.

主な機能

• Simulates common MCP vulnerabilities for educational purposes.
• Provides example exploits for each vulnerability.
• Includes documentation on identification, impact, and mitigation strategies.
• Supports context manipulation, prompt injection, and access control bypass scenarios.
• Allows for exploration of model chain attacks and response manipulation.
• 2 GitHub stars

ユースケース

• Training security researchers in AI/ML security.
• Testing and validating security tools and techniques for AI/ML systems.
• Educating developers about potential vulnerabilities in model serving.