CUAHarm/CUAHarm

CUAHarm数据集是一个用于评估计算机使用代理（CUAs）安全风险的基准，包含了104个由专家编写的现实滥用风险场景。这些场景涵盖了凭证盗窃、权限提升、网络利用、系统破坏、数据篡改、逃避取证和工具利用等七种恶意目标类别。数据集提供了两种配置文件：computer_use_tasks.csv包含需要直接系统交互的计算机使用任务，而common_malicious_prompts.csv包含了用于评估常见恶意活动的额外提示。

The CUAHarm dataset is a benchmark designed to evaluate the safety risks of Computer-Using Agents (CUAs) and consists of 104 expert-written realistic misuse scenarios. These scenarios cover seven categories of malicious objectives: credential theft, privilege escalation, network exploitation, system disruption, data tampering, forensic evasion, and tool utilization. The dataset includes two configuration files: computer_use_tasks.csv with detailed computer use tasks involving direct system interaction, and common_malicious_prompts.csv with additional prompts for evaluating agents against common malicious activities.