SafeMCP

该数据集名为SafeMCP，是一个受控框架，旨在通过模拟不同的提示注入攻击策略来检查MCP驱动代理系统的安全问题。该数据集包含了多种攻击方法以及用于评估MCP驱动代理系统安全性的各项指标。其规模覆盖了20个代理任务，横跨多个模型家族。该数据集的任务是评估MCP驱动代理系统中的安全风险及防御机制。The dataset, named SafeMCP, is a controlled framework designed to examine security vulnerabilities in MCP-powered agent systems by simulating various prompt injection attack strategies. This dataset includes multiple attack methods and a comprehensive set of metrics for evaluating the security of MCP-powered agent systems. It covers 20 agent tasks across multiple model families. The core task of this dataset is to assess security risks and defense mechanisms within MCP-powered agent systems.