研究人员使用Anthropic的Claude AI模型对25个大型开源项目进行了自动化攻击面分析1。这些项目涵盖超过500万行代码，使用C、C++、Java、PHP和Python等多种语言1。分析结果发现493个安全问题1，其中包括93个不必要的预身份验证接口、37个身份验证前执行的代码，以及61个弱化安全控制的配置组合1。
在162个安全敏感的配置组合中，94个无声关闭了安全控制，68个存在机制不一致1。Claude通过分析默认配置、配置交互和代码可达性等方法，识别出不必要暴露的代码路径。其中最显著的发现是在OpenSSH中发现了约4500行ML-DSA-44签名解析代码不必要地暴露给未授权客户端，占该项目预身份验证代码的约16%1。针对这一漏洞，OpenSSH 10.5版本修复了第一个问题，10.6版本在多轮迭代后完全修复，移除了大约五分之一的未授权客户端可达的代码1。
研究还发现相同的漏洞模式独立出现在四个无关项目中，均涉及主机名DNS反向查询配置与拒绝列表的交互导致列表失效开启1。这表明AI模型在减少代码攻击面、提高已部署软件安全性方面具有实际应用价值1。
Anthropic's Claude AI model has been employed to conduct automated attack surface analysis across 25 large open-source projects containing over 5 million lines of code, uncovering 493 security issues 1. The analysis examined codebases written in C, C++, Java, PHP, and Python, identifying unnecessary code paths exposed to potential attackers through three analytical methods: default configuration review, configuration interactions, and code reachability assessment 1.
The findings revealed significant security concerns across multiple vulnerability categories 1. Claude identified 93 unnecessary pre-authentication interfaces, 37 instances of code executing before authentication, and 61 configuration combinations that weakened security controls 1. Among 162 security-sensitive configuration combinations analyzed, 94 silently disabled security controls while 68 exhibited mechanistic inconsistencies 1. The same vulnerability pattern—involving hostname DNS reverse query configuration interacting with deny lists to disable filtering—appeared independently across four unrelated projects 1.
OpenSSH presented a particularly striking case study 1. Researchers discovered approximately 4,500 lines of ML-DSA-44 signature parsing code unnecessarily exposed to unauthenticated clients, representing roughly 16 percent of OpenSSH's pre-authentication code 1. In response, OpenSSH version 10.5 patched the initial vulnerability, with version 10.6 achieving complete remediation after multiple development iterations by removing approximately one-fifth of the unauthenticated client-reachable code 1. These results demonstrate the practical value of AI models in enhancing the security posture of widely deployed software by systematically reducing code attack surfaces 1.
评论
还没有评论，欢迎留下第一条。