AI Security AI安全 1d ago Updated 1d ago 更新于 1天前 46

Researchers Use Claude to Port Pre-Auth RCE Exploit From One PLC Model to Another 研究人员利用 Claude 将预认证 RCE 漏洞利用从一种 PLC 模型移植到另一种

Forescout Research/Vedere Labs used Anthropic's Claude to port a working pre-auth RCE exploit (CVE-2021-31886) from a WAGO 750-852 PLC to a different model, the 750-831, executing ARM shellcode on live hardware The exploit targets a stack-based buffer overflow in the Nucleus FTP server's USER command handling (CVSS 9.8) over unauthenticated TCP port 21, with no firmware updates available from the vendor Claude adapted the attack by switching from a USER/QUIT sequence to USER/CWD with an omitted 研究人员使用Claude AI成功将针对WAGO PLC的预认证RCE漏洞利用从750-852型号移植到750-831型号,在真实硬件上执行了ARM shellcode 利用CVE-2021-31886(Nucleus FTP服务器栈溢出,CVSS 9.8),通过修改FTP命令序列绕过缓冲区清零机制实现代码执行 整个开发过程耗时8小时32分钟,API费用达535.74美元,需要研究人员持续引导和干预 尝试将漏洞利用扩展为C2植入物时意外写入flash内存区域,永久损坏了目标PLC 研究揭示了AI降低ICS漏洞利用技术门槛的现实风险,与NSA/CISA等机构发布的AI生成利用脚本警告相呼应

65
Hot 热度
70
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Forescout Research/Vedere Labs used Anthropic's Claude to port a working pre-auth RCE exploit (CVE-2021-31886) from a WAGO 750-852 PLC to a different model, the 750-831, executing ARM shellcode on live hardware
  • The exploit targets a stack-based buffer overflow in the Nucleus FTP server's USER command handling (CVSS 9.8) over unauthenticated TCP port 21, with no firmware updates available from the vendor
  • Claude adapted the attack by switching from a USER/QUIT sequence to USER/CWD with an omitted CRLF terminator to bypass buffer zeroing, achieving functional RCE payloads (ICMP and UDP "PWNED") within 12 minutes of initial code execution
  • The session cost $535.74 in API usage over 8 hours 32 minutes, and a subsequent attempt to build a C2 implant permanently bricked the PLC by writing to flash-mapped memory
  • This follows a joint NSA/CISA/FBI/DOE/EPA advisory warning of active AI-generated exploitation scripts targeting internet-exposed Siemens PLCs, signaling a broader shift in ICS threat actor capabilities

Why It Matters

This research demonstrates that AI coding agents can successfully port complex zero-day-class exploits across different hardware platforms with minimal human intervention, dramatically lowering the barrier to attacking operational technology infrastructure. For defenders, it underscores that vulnerabilities disclosed years ago in widely deployed PLCs remain actively exploitable through AI-assisted means, and that the risk calculus around unpatched OT assets must be fundamentally reconsidered.

Technical Details

  • Vulnerability: CVE-2021-31886, a stack-based buffer overflow in the Nucleus FTP server's handling of the USER command on WAGO PLCs, rated CVSS 9.8, accessible pre-authentication over TCP port 21
  • AI Agent Setup: Claude Code (starting with Sonnet 4.6, upgraded to Opus 4.6) was given access to a terminal, Ghidra reverse-engineering tool, the target PLC firmware binary, and the live hardware target, running interactive sessions with researcher steering
  • Exploit Adaptation: The original 750-852 exploit used a USER/QUIT sequence; Claude discovered that switching to USER/CWD and omitting the CRLF terminator prevented the FTP processing path from zeroing the 256-byte attacker-controlled buffer, allowing shellcode to survive long enough to execute
  • Payload Execution: Once RCE was established, Claude generated two functional payloads in 12 minutes — one sending ICMP echo requests and another sending a UDP packet containing "PWNED" — running in the Ethernet receive callback context
  • Additional Findings: Claude flagged a potential second vulnerability in the FTP command extraction loop distinct from CVE-2021-31886, which manual review suggested may be a previously unidentified flaw (no CVE assigned); the exploit is absent from public repositories (Exploit-DB, Packet Storm, GitHub)
  • Bricking Incident: A follow-up session attempting to extend the exploit into a C2 implant wrote to a flash-mapped memory region, permanently damaging the physical PLC

Industry Insight

  • Organizations running WAGO PLCs (and other Nucleus V1 RTOS-based controllers) with internet-exposed or poorly segmented FTP services face immediate, actionable risk; until firmware updates exist, network segmentation and port 21 blocking are the only mitigations
  • The AI-assisted exploit porting demonstrates that threat actors no longer need deep reverse-engineering expertise to adapt known vulnerabilities to new targets, validating government warnings about AI-amplified ICS threats and accelerating the need for proactive vulnerability management in OT environments
  • The bricking incident highlights a critical operational risk: AI agents operating on physical systems can cause irreversible hardware damage, suggesting that any use of autonomous AI agents in OT contexts requires strict sandboxing, human-in-the-loop safeguards, and clear boundaries between virtual and physical system access

TL;DR

  • 研究人员使用Claude AI成功将针对WAGO PLC的预认证RCE漏洞利用从750-852型号移植到750-831型号,在真实硬件上执行了ARM shellcode
  • 利用CVE-2021-31886(Nucleus FTP服务器栈溢出,CVSS 9.8),通过修改FTP命令序列绕过缓冲区清零机制实现代码执行
  • 整个开发过程耗时8小时32分钟,API费用达535.74美元,需要研究人员持续引导和干预
  • 尝试将漏洞利用扩展为C2植入物时意外写入flash内存区域,永久损坏了目标PLC
  • 研究揭示了AI降低ICS漏洞利用技术门槛的现实风险,与NSA/CISA等机构发布的AI生成利用脚本警告相呼应

为什么值得看

这项研究首次实证展示了AI如何被用于自动化移植工业控制系统漏洞利用,为网络安全行业提供了关于AI赋能攻击者的具体案例。对于OT安全从业者而言,这重新定义了ICS漏洞的威胁等级和应急响应优先级。

技术解析

  • 漏洞利用基于CVE-2021-31886,针对Nucleus FTP服务器的USER命令处理中的栈溢出,通过修改FTP命令序列(USER+CWD替代USER+QUIT)并省略CRLF终止符来绕过缓冲区清零机制
  • 研究团队向Claude Code提供了现有漏洞利用代码、目标固件二进制文件和物理PLC设备,AI通过Ghidra逆向工程工具和终端交互完成漏洞适配,初始使用Claude Sonnet 4.6,后升级到Claude Opus 4.6
  • 成功执行后,AI在12分钟内从NOP shellcode演进为两个功能性payload:ICMP回显请求和包含"PWNED"字符串的UDP数据包,漏洞利用运行在以太网接收回调上下文
  • 漏洞利用运行在以太网接收回调上下文,当前演示能力仅限于发送网络数据包,尚未实现持久化C2植入

行业启示

  • AI正在显著降低工业控制系统漏洞利用的技术门槛和时间成本,威胁行为者可能快速将已知漏洞转化为针对特定目标的定制化攻击工具,组织需重新评估ICS/OT环境中的漏洞优先级
  • 缺乏厂商补丁支持的遗留系统(如基于Nucleus V1 RTOS的WAGO PLC)面临更高风险,安全团队应加强网络分段、监控高风险协议流量并建立针对AI生成利用脚本的检测能力
  • 研究意外发现了一个潜在的独立漏洞(FTP命令提取循环中的bug),表明AI辅助分析可能揭示新的安全缺陷,但也凸显了AI操作物理系统时的不可控风险

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Claude Claude Security 安全 Research 科学研究 LLM 大模型