VPI-Bench/vpi-bench

VPI-Bench是一个用于评估计算机使用代理和浏览器使用代理在视觉提示注入攻击下鲁棒性的基准数据集，包含测试用例和网页平台。测试用例以JSON文件形式存储，包含唯一的ID、URL、用户提示、攻击者目标、平台类型和环境配置命令。网页平台已在GitHub上部署。数据集旨在提供现实的对抗性威胁场景，不包含任何真实的用户数据或敏感个人信息。

VPI-Bench is a benchmark dataset for evaluating the robustness of computer-use and browser-use agents under visual prompt injection attacks, containing testcases and web platforms. Testcases are stored in JSON format, including a unique ID, URL, user prompt, attacker objective, platform type, and environment setup commands. Web platforms have been deployed on GitHub. The dataset aims to provide realistic adversarial threat scenarios and does not contain any real user data or sensitive personal information.