WARD-Base
收藏资源简介:
WARD-Base是由新加坡国立大学等机构构建的大规模网页代理安全防御数据集,旨在提升对提示注入攻击的检测能力。该数据集包含约17.7万条样本,覆盖719个高流量网站和10个模拟高风险平台,数据来源于真实网页探索和模拟用户生成内容,通过双分支管道构建恶意与良性样本。数据集构建过程采用两阶段方法:首先收集网页HTML与截图作为基础数据,随后通过视觉语言模型生成符合上下文的提示注入内容。该数据集主要应用于网页代理安全领域,用于训练和评估防御模型,以解决开放网络环境中提示注入攻击导致的代理决策操纵和安全漏洞问题。
WARD-Base is a large-scale web proxy security defense dataset developed by institutions including the National University of Singapore, aiming to enhance the detection capability against prompt injection attacks. This dataset contains approximately 177,000 samples, covering 719 high-traffic websites and 10 simulated high-risk platforms. The data is sourced from real web crawls and simulated user-generated content, with malicious and benign samples constructed via a two-branch pipeline. The dataset construction adopts a two-stage approach: first, web HTML and screenshots are collected as basic data, followed by the generation of context-aware prompt injection content via vision-language models. This dataset is primarily applied in the field of web proxy security, used for training and evaluating defense models to address proxy decision manipulation and security vulnerabilities caused by prompt injection attacks in open network environments.
数据集概述:WARD (WebAgent Adversarial Robust Defense)
- 数据集名称:WARD (Adversarially Robust Defense of Web Agents Against Prompt Injections)
- 项目主页:https://caothientri2001vn.github.io/WARD-WebAgent/
- 论文地址:https://arxiv.org/abs/2605.15030
- 模型仓库(Hugging Face):
- 数据发布状态:评估和训练数据即将发布。
数据集核心任务
- 领域:Web Agent安全防御
- 任务类型:多模态提示注入检测(Prompt Injection Detection)
- 输入:用户意图任务 + 网页预处理HTML文本 + 网页截图
- 输出:一个结构化的JSON,包含:
reasoning:基于证据的推理分析attack_goal:攻击目标描述(若无攻击则为"none")injection_location:注入位置("html"、"screenshot"、"both"或"none")label:样本标签("malicious"或"benign")
模型与使用
- 模型类型:多模态守卫检查点(Multimodal Guard Checkpoints)
- 可用模型:
WARD-0.8b和WARD-2b - 运行环境:依赖
torch、pillow、transformers - 推理示例:提供了完整的
transformers代码示例,包括加载模型、处理器、构造对话模板、生成推理结果。




