weijianzhg/email-guard-prompt-injection-tt

Email Guard Prompt-Injection Dataset是一个公开安全的指令调优数据集，包含800行数据，用于训练或评估小型电子邮件防护模型。任务是对不受信任的电子邮件文本进行摘要，并标记邮件是否包含针对AI系统的提示注入指令。数据集结构为扁平JSONL格式，每行包含input（电子邮件主题和正文）和output（JSON字符串，包含摘要和攻击标记）字段。数据来源为合成数据，其中400行为提示攻击行，400行为非攻击行，攻击输出不复制注入的有效载荷，而是将其摘要为提示注入尝试。数据集旨在用于微调或测试一个位于大型电子邮件处理工作流前端的防护模型，适用于电子邮件分类、支持票证接收、文档摄取等场景。数据集为英文，仅限于AI目标提示注入攻击检测，不涵盖其他类型的攻击如钓鱼、垃圾邮件或恶意软件。Email Guard Prompt-Injection Dataset is a public-safe 800-row instruction-tuning dataset for training or evaluating a small email guard model. The task is to summarize untrusted email text and flag whether the email contains prompt-injection instructions aimed at an AI system. The dataset is in flat JSONL format with input and output fields, where output is a JSON string containing summary and prompt_attack. It is synthetic-only, with 400 prompt-attack rows and 400 non-attack rows, and attack outputs summarize the injection as a prompt-injection attempt without copying the payload. The intended use is to fine-tune or test a guard model that sits in front of a larger email-processing workflow for applications like email triage, support-ticket intake, and document ingestion. The dataset is English-only and focuses narrowly on AI-targeted prompt injection, excluding other attack types such as phishing, spam, or malware.