SEC-bench
收藏资源简介:
SEC-bench是一个自动化的基准测试框架,用于评估大型语言模型(LLM)代理在真实软件安全任务上的能力。该数据集包含了200个真实世界的CVE实例,每个实例都经过验证并具有可重现的漏洞利用(PoC)和经过验证的安全补丁。数据集的创建过程采用了多代理支撑结构,自动构建带有测试用例的代码仓库,在隔离环境中重现漏洞,并生成金标补丁以进行可靠评估。数据集的访问地址为https://hf.co/datasets/SEC-bench/SEC-bench。数据集旨在解决LLM代理在真实世界软件安全任务上的性能评估问题,通过实现两个关键的软件安全任务(PoC生成和漏洞补丁)来评估LLM代理的能力。
SEC-bench is an automated benchmarking framework designed to evaluate the capabilities of Large Language Model (LLM) Agents on real-world software security tasks. This dataset contains 200 real-world CVE instances, each of which has been validated, with reproducible Proof-of-Concept (PoC) exploits and verified security patches. The dataset was constructed using a multi-agent supported pipeline, which automatically builds code repositories with test cases, reproduces vulnerabilities in isolated environments, and generates gold-standard patches for reliable evaluation. The dataset is accessible at https://hf.co/datasets/SEC-bench/SEC-bench. This dataset aims to address the performance evaluation gap of LLM Agents on real-world software security tasks, by evaluating their capabilities through two core software security tasks: Proof-of-Concept generation and vulnerability patching.




