AI helped port an exploit between PLC models. Success still took hours and more than $500

Researchers at Forescout Vedere Labs tested whether a large language model could help port a working exploit between two industrial WAGO PLC models. Using Claude Code, they eventually achieved working code execution, but the experiment also showed how far current systems remain from fully autonomous exploit development.

From one PLC model to another

The starting point was an existing exploit for the WAGO 750-852 based on CVE-2021-31886, a pre-authentication buffer overflow in the Nucleus FTP server that can allow arbitrary ARM code execution. Claude was asked to adapt it to the related but different WAGO 750-831.

The model had access to a terminal, reference files, the Ghidra reverse-engineering tool and the physical target device. Claude confirmed the vulnerability and quickly generated a payload that crashed the PLC, but moving from a crash to controlled code execution was much harder.

The AI still needed expert guidance

Early attempts sent the model down incorrect paths. Researchers had to redirect the analysis and provide additional technical context. Progress improved after the team switched from Claude Sonnet 4.6 to Claude Opus 4.6 and instructed the model to ask for help whenever it was uncertain about a firmware detail.

Once Claude solved the issue that caused injected code to be erased before execution, iteration accelerated sharply. SecurityWeek reports that the model produced two working payloads within 12 minutes.

More than eight hours and over $500 for the final RCE stage

Forescout said the final stage of remote-code-execution development alone consumed more than $500 in API usage during a session lasting more than eight hours. That matters because the result is not evidence of cheap one-click autonomous hacking. It is evidence that a skilled researcher can use AI as a fast analysis and iteration partner.

Not every iteration succeeded

In a separate follow-up session aimed at building a command-and-control implant, one payload wrote to flash-mapped memory and permanently bricked the PLC. That failure is another reminder that current models still require supervision and can take costly wrong turns.

What businesses should take from it

The practical conclusion is not that AI can now independently hack factories. Specialist knowledge is still required. The more important change is that AI may shorten parts of the analysis and exploit-adaptation workflow. For organisations running OT and PLC environments, that increases the value of network segmentation, reduced device exposure, timely patching and monitoring for unusual traffic.

Sources

SecurityWeek: Experiment: Porting a PLC Exploit With AI Takes Hours and Hundreds of Dollars Forescout: R4IoT research report ? background on WAGO PLC exploitation and CVE-2021-31886