CySecBench
收藏资源简介:
CySecBench是由瑞典皇家理工学院网络系统安全组创建的一个专注于网络安全领域的数据集,旨在评估大型语言模型在生成恶意代码时的抗越狱能力。该数据集包含12662条提示,分为10个不同的攻击类型类别,涵盖了从云攻击到物联网攻击的广泛网络安全场景。数据集的生成过程使用了OpenAI的GPT模型,通过生成和过滤恶意提示来确保数据的质量和针对性。CySecBench的应用领域主要集中在网络安全研究,特别是评估和提升大型语言模型在生成恶意代码时的安全性。该数据集的发布为研究人员提供了一个标准化的工具,用于评估和改进语言模型在网络安全领域的表现。
CySecBench is a cybersecurity-focused dataset developed by the Cybersecurity and Network Systems Group at KTH Royal Institute of Technology. It is designed to evaluate the jailbreak resistance of large language models (LLMs) when generating malicious code. This dataset comprises 12,662 prompts, categorized into 10 distinct attack type classes, covering a broad spectrum of cybersecurity scenarios ranging from cloud attacks to IoT attacks. The dataset was constructed using OpenAI's GPT models, with malicious prompts generated and filtered to ensure data quality and targeted specificity. The primary application areas of CySecBench are concentrated in cybersecurity research, specifically for assessing and enhancing the safety of LLMs during malicious code generation. The release of CySecBench provides researchers with a standardized tool to evaluate and improve the performance of language models in the cybersecurity domain.




