소개
SecBench is a comprehensive security benchmark and interactive playground designed to rigorously test the robustness of Model Context Protocols (MCPs). It facilitates the identification of vulnerabilities by simulating various attack scenarios, including malicious server interactions, Man-in-the-Middle attacks, DNS rebinding, and specific CVE exploits. The tool supports automated testing against popular LLM agents like OpenAI, Claude, and Cursor, and offers a flexible client to connect with both normal and malicious MCP servers, making it an essential resource for developers and researchers focused on securing AI agent interactions.
주요 기능
- Automated security testing for OpenAI, Claude, and Cursor LLMs
- Simulates various attacks including server name squatting, Man-in-the-Middle, DNS rebinding, and CVE exploitation
- Customizable client for interacting with MCP hosts and servers
- Includes pre-configured malicious servers for diverse vulnerability testing scenarios
- Designed for extensibility to other LLM models (e.g., Deepseek, Llama, QWen)
- 6 GitHub stars
사용 사례
- Evaluating LLM agents for susceptibility to security vulnerabilities and attacks
- Developing and testing custom MCP clients and servers in a controlled, adversarial environment
- Benchmarking the security of Model Context Protocol (MCP) implementations
Advertisement