Experiment: Porting a PLC Exploit With AI Takes Hours and Hundreds of Dollars
Forescout's Vedere Labs used Anthropic's Claude to port a working RCE exploit from a WAGO 750-852 PLC to a WAGO 750-831 PLC, demonstrating AI-assisted exploit adaptation for industrial control systems The experiment was based on CVE-2021-31886, a pre-authentication buffer overflow in the Nucleus FTP server allowing arbitrary ARM shellcode execution Claude Code was given access to a terminal, reference files, Ghidra, and the physical target device, confirming the vulnerability through live probin
Analysis
TL;DR
- Forescout's Vedere Labs used Anthropic's Claude to port a working RCE exploit from a WAGO 750-852 PLC to a WAGO 750-831 PLC, demonstrating AI-assisted exploit adaptation for industrial control systems
- The experiment was based on CVE-2021-31886, a pre-authentication buffer overflow in the Nucleus FTP server allowing arbitrary ARM shellcode execution
- Claude Code was given access to a terminal, reference files, Ghidra, and the physical target device, confirming the vulnerability through live probing and static firmware analysis
- Progress required switching from Claude Sonnet 4.6 to Claude Opus 4.6 and instructing the model to ask for help when uncertain about firmware details
- The final RCE development stage consumed over $500 in API costs across more than eight hours, and a follow-up C2 implant attempt permanently bricked the PLC by writing to flash memory
Why It Matters
This experiment highlights the growing capability of AI models to assist in adapting exploits across similar hardware targets in OT/ICS environments, a concern amplified by recent attacks on water sector PLCs. While current costs and human oversight requirements remain high, the pattern of rapid iteration after initial breakthrough suggests that marginal costs could decrease significantly as models improve, enabling attackers to scale exploit adaptation across many targets simultaneously.
Technical Details
- The starting exploit targeted CVE-2021-31886, a pre-authentication buffer overflow in the Nucleus FTP server of WAGO 750-852 PLCs, enabling unauthenticated ARM shellcode execution
- Claude Code was equipped with a terminal, reference files, Ghidra reverse-engineering tool, and physical access to the target WAGO 750-831 PLC
- Early attempts failed because injected code was being erased before execution; the breakthrough came when Claude identified this firmware-level interference and adjusted its payload delivery approach
- After the breakthrough, Claude generated two separate working payloads within 12 minutes, demonstrating rapid iterative capability once the core obstacle was understood
- A subsequent session attempting to build a command-and-control implant resulted in a payload that wrote to flash memory, permanently bricking the device
Industry Insight
- The "hard start, fast follow" pattern observed—where initial exploitation is time-consuming but subsequent iterations are rapid—suggests that as AI models improve, the bottleneck will shift from discovery to scaling across multiple targets, making mass exploit adaptation a realistic threat
- The $500+ cost and 8+ hour timeline for a single exploit port, combined with the need for expert redirection, indicates that AI-assisted OT exploitation is not yet cost-effective for individual attackers but could become viable as intervention requirements decrease
- Organizations should treat AI-augmented exploit development as an emerging threat vector for ICS/OT systems and prioritize hardening PLC firmware, implementing network segmentation, and monitoring for anomalous exploitation patterns consistent with AI-assisted attack workflows
Disclaimer: The above content is generated by AI and is for reference only.