CyberSecEval
A cybersecurity AI capability evaluation benchmark introduced by Meta. It assesses models across two dimensions: the quality of defensive knowledge outputs and the suppression of potentially harmful capabilities.
Timeline (last 90 days)
CyberSecEval introduced metrics such as "cyberattack assistance rate" in an attempt to quantify the extent to which a model assists potentially malicious behavior.
CyberSecEval is a security benchmark released by Meta, used to evaluate model performance on code vulnerability identification, CTF problem-solving, and security Q&A.
CyberGym是专门评估AI模型在网络安全和代码安全审查领域能力的基准框架，涵盖漏洞检测、恶意代码识别、安全配置审计、依赖项风险评估等维度
网络安全能力基准测试的类似框架包括DARPA的Cyber Grand Challenge和Meta的PurpleLlama CyberSecEval
All Facts (4)
CyberSecEval introduced metrics such as "cyberattack assistance rate" in an attempt to quantify the extent to which a model assists potentially malicious behavior.
50%UnverifiedCyberSecEval is a security benchmark released by Meta, used to evaluate model performance on code vulnerability identification, CTF problem-solving, and security Q&A.
50%UnverifiedCyberGym是专门评估AI模型在网络安全和代码安全审查领域能力的基准框架，涵盖漏洞检测、恶意代码识别、安全配置审计、依赖项风险评估等维度
50%Unverified网络安全能力基准测试的类似框架包括DARPA的Cyber Grand Challenge和Meta的PurpleLlama CyberSecEval
50%