Researchers at Forescout's Vedere Labs recently conducted an experiment using Anthropic's Claude AI to port a remote code execution exploit from one WAGO programmable logic controller to another. The experiment aimed to adapt an existing exploit, initially developed for the WAGO 750-852 PLC, to the WAGO 750-831 model. The original exploit was based on CVE-2021-31886, a vulnerability in the Nucleus FTP server that allowed attackers to execute arbitrary ARM shellcode without authentication. The process took several hours, required extensive researcher oversight, and resulted in API costs totaling hundreds of dollars.
The team used Claude Code in conjunction with tools like Ghidra for reverse engineering and direct access to the target device. The AI confirmed the presence of the vulnerability through live probing and firmware analysis but initially struggled to achieve controlled code execution. Researchers had to intervene multiple times to correct the AI's direction and provide technical context. Progress accelerated when the team switched to Claude Opus 4.6 and instructed the AI to seek help when uncertain about firmware details. The AI eventually succeeded in producing two working payloads within minutes once it overcame an issue with the injected code being erased before execution.
However, attempts to develop a more complex command-and-control implant were less successful. During this phase, one payload inadvertently wrote to a region of the PLC's flash memory, causing permanent damage to the device. The researchers noted that while AI-assisted porting is currently more time-consuming and expensive than manual methods, the potential for reducing the marginal cost of exploiting multiple targets simultaneously could be significant as AI technology advances.