Security researchers have demonstrated that generative AI can rapidly adapt known vulnerabilities across different programmable logic controller (PLC) models, effectively dismantling a long-standing assumption in industrial cybersecurity. Forescout Research’s Vedere Labs successfully used Anthropic’s Claude to port a pre-authentication remote code execution (RCE) exploit between two distinct WAGO PLCs, ultimately executing custom ARM shellcode on live hardware.
The proof-of-concept targeted CVE-2021-31886, a stack-based buffer overflow in the Nucleus FTP server’s handling of the USER command. Historically, adapting such an exploit to different hardware architectures or firmware variants required weeks of manual reverse engineering. The Vedere Labs experiment shows that large language models can compress this process into a near-instantaneous pipeline, autonomously managing payload translation, memory offset calibration, and environmental constraints.
This capability neutralizes the implicit security buffer that operational technology (OT) environments have historically relied upon: hardware fragmentation and extended patch cycles. When AI can systematically bridge architectural gaps, a single pre-authentication flaw transforms into a scalable, multi-vendor threat. The finding underscores that traditional industrial defenses, which often depend on obscurity and slow maintenance windows, are no longer viable against machine-speed attack development.
In response, security practitioners are being urged to abandon legacy reliance on fragmented ecosystems and adopt proactive, automated defenses. Experts recommend implementing continuous, real-time firmware inventory tracking, accelerating patch deployment or deploying compensating controls for known vulnerabilities, and permanently disabling unauthenticated legacy services like FTP and Telnet. Zero-trust network segmentation remains critical to restrict lateral movement and contain unauthenticated access before an exploit can propagate.
The dual-use nature of the technology means defensive operations must evolve at an equivalent velocity. Integrating AI-assisted threat emulation, automated code auditing, and continuous vulnerability triage can help security teams keep pace. However, deployment in legacy OT environments introduces significant operational and compliance hurdles. Industry leaders are currently weighing how to safely validate AI-driven patches without risking safety-critical failures, while also navigating data privacy regulations when routing sensitive telemetry to third-party commercial AI models. Questions also persist regarding vendor accountability for end-of-life industrial components in an AI-accelerated threat landscape.
The demonstration marks a definitive inflection point for industrial cybersecurity. As AI eliminates the friction of cross-architecture exploit development, organizations must prioritize automated, continuous validation and strict network isolation to maintain operational resilience.
安全研究人員已證明，生成式AI能夠迅速將已知漏洞移植至不同型號的可程式邏輯控制器（PLC），有效打破工業網絡安全領域長期以來的一項假設。Forescout Research旗下的Vedere Labs成功利用Anthropic的Claude，將一個免認證的Remote Code Execution (RCE) exploit移植至兩款不同的WAGO PLC之間，最終在實體硬件上成功執行自訂的ARM shellcode。
該概念驗證針對CVE-2021-31886，此為Nucleus FTP伺服器處理USER指令時出現的stack-based buffer overflow。過往，將此類exploit移植至不同硬件架構或韌體版本，往往需要數週時間進行手動reverse engineering。Vedere Labs的實驗顯示，大型語言模型能夠將此過程壓縮至近乎即時的pipeline，自主處理payload轉換、記憶體偏移量校準及環境限制。
此能力抵消了操作技術（OT）環境長期以來依賴的隱性安全緩衝：硬件碎片化及冗長的修補週期。當AI能夠系統性地彌合架構差異時，單一免認證漏洞便會演變為具擴展性、跨供應商的威脅。此發現凸顯，傳統工業防禦往往依賴隱蔽性及緩慢的維護視窗，面對以機器速度開發的攻擊已不再可行。
為此，業界敦促網絡安全專業人員放棄過往對碎片化生態系統的依賴，轉而採取主動及自動化的防禦措施。專家建議實施持續且即時的韌體資產追蹤、加快修補程式部署，或針對已知漏洞部署補償性控制措施，並永久停用FTP及Telnet等免認證的舊式服務。零信任網絡分段仍是限制橫向移動的關鍵，能在exploit擴散前有效封鎖未經認證的存取。
該技術的雙重用途特性意味著防禦行動必須以同等速度演進。整合AI輔助的威脅模擬、自動化程式碼審計及持續的漏洞分類，可協助安全團隊跟上步伐。然而，在舊式OT環境中部署將帶來重大的營運及合規障礙。業界領袖目前正權衡如何安全地驗證AI驅動的修補程式，以免引發涉及安全關鍵的故障；同時在將敏感遙測數據傳送至第三方商業AI模型時，亦需應對數據私隱法規。在AI加速威脅的環境下，供應商對已終止支援工業組件的問責問題亦持續引發關注。
是次演示標誌著工業網絡安全的明確轉折點。隨著AI消除跨架構exploit開發的阻力，機構必須優先推行自動化、持續驗證及嚴格的網絡隔離，以維持營運韌性。