Bordair Multimodal Prompt Injection Dataset
收藏资源简介:
Bordair多模态提示注入数据集包含62,063个标记样本(38,304个攻击样本和23,759个良性样本),覆盖跨模态、多轮、对抗性后缀、越狱模板、间接注入、工具操纵和规避攻击等多种攻击类型。数据集旨在训练和评估提示注入检测器,所有样本均标记(`expected_detection: true/false`),来源可追溯到同行评审论文或行业研究,并结构化以便直接用于二元分类器。
The Bordair Multimodal Prompt Injection Dataset comprises 62,063 labeled samples, consisting of 38,304 adversarial samples and 23,759 benign samples. It encompasses multiple attack categories including cross-modal, multi-turn, adversarial suffix, jailbreak template, indirect injection, tool manipulation, and evasion attacks. This dataset is intended for the training and evaluation of prompt injection detectors, with every sample annotated with the tag `expected_detection: true/false`. All sample sources are traceable to peer-reviewed academic papers or industrial research, and the dataset is structured to enable direct utilization in binary classifiers.




