CaiZhiTech/Evaluation-Dataset-of-AI-Agent-Security-Guardrails

DKnownAI代理安全评估数据集是一个用于评估AI代理安全性的数据集，特别关注对抗性输入（提示）及其被安全防护栏分类为阻止或允许的情况。数据集包含两个主要字段：text表示对抗性输入，action表示人工标注的标签（阻止或允许）。该数据集适用于文本分类任务，使用英语，涉及安全、提示注入、越狱、代理安全和防护栏评估等标签。数据集规模在1K到10K之间。

The DKnownAI Agent Security Evaluation Dataset is designed for evaluating the security of AI agents, specifically focusing on adversarial inputs (prompts) and their classification as blocked or allowed by security guardrails. The dataset includes two main fields: text representing the adversarial input and action representing the human-annotated label (blocked or allowed). It is suitable for text-classification tasks, uses English, and involves tags such as security, prompt-injection, jailbreak, agent-safety, and guardrail-evaluation. The dataset size ranges between 1K and 10K.