做完这个工作后发现,大模型在找漏洞中是有用武之地的,但并不能代替我们传统的程序分析,不管是静态的照旧动态的分析,它们是相对精准的,并且是大模型不善于的。但是,大模型有能力非常好地总结和阅读代码,该举动恰好是黑客在找漏洞过程中所用的方法,而传统方法或工具实在并没有真正地明白代码。因此,现有大部分自动化工具都基于非常明确的规则来找漏洞,而没有通过语义明白和推理实现,最终做了一个GPTScan工作。 重点4: GPTScan基本思绪就是模拟安全审计人员他们的思维过程,把这个过程分成了Planning、Reasoning和Validating三个部分,重要的是想安全专家一样明白代码(安全知识)。
GPTScan proposes the first LLM-based static vulnerability detection pipeline,combining LLM with program analysis to achieve improved accuracy (~70% F1-score). It also identifies 9 new vulnerabilities missed by human auditors.
由于直播视频的部分内容杂音实在太大,非常抱歉,只能推荐读者阅读原文。下面给出这篇文章的结论:
This paper conducts a comprehensive study on the repair capabilities of LLMCs in various repair scenarios under the NMT fine-tuning paradigm. Our results show that even without any post-processing strategies, LLMCs can already achieve excellent results, and surpass many previous APR works. Importantly, we present some practical guidelines on how to choose different designs to better exploit the repair capability of LLMCs, and show how they can repair complex defects. We also analyze and discuss some limitations found during the evaluation and point out future directions. Furthermore, our results on various benchmarks can serve as the baselines for subsequent works with reference. In conclusion, LLMC-based APR has great potential for practical use, and more efforts are needed to promote LLM4APR research in the future.
论文下载地址: