Collinear AI发布CWE-bench v2:100个漏洞修复任务覆盖91类CWE
Collinear AI于11月3日发布CWE-bench v2,面向公众开放100个任务,用于评估AI模型在真实开源仓库中查找并修复安全漏洞的能力。任务覆盖91种CWE漏洞类别,评分完全程序化,结果可复现且无需额外成本。
在已评估的11个模型中,Gemini 4 Argon以67.8%的pass@1和81.0%的pass@4领先,Claude Opus 5.5以67.3%和78.0%紧随其后。95个任务至少被一个模型解决,但领先模型仍有19%的任务在所有尝试中失败。
成本方面,DeepSeek V4.1 Flash(pass@1 57.0%,每次运行US$0.04)、Claude Opus 5.5(每次运行US$1.83)和Gemini 4 Argon(每次运行US$3.92)构成Pareto前沿。每个任务都是自包含的Harbor任务,可通过Hugging Face仓库直接运行。数据集设有访问门槛,需先在数据集页面接受访问条件;另有保留集可用于进一步评估,联系cyber@collinear.ai获取。任务还标记为rl-environment,可提供奖励信号用于训练防御型智能体。
- 2026-11-03 08:00 | Huggingface:CWE-bench v2: More coverage, quality, and transparency阅读原文
CWE-bench is designed to test the limits of frontier cyber capabilities. In our continued pursuit to push the horizon at Collinear AI, CWE-bench v2 opens up a public set of 100 tasks that anyone can run and evaluate their models on. The setup is the same as before. An agent gets a checkout of a real...