More than $500 in API usage across a session of more than eight hours works out to roughly $60 an hour in inference spend [13][15], and that covers only the final stage of the port. A researcher sat with it the whole way, supplying firmware context and pulling the model off incorrect leads [7].
No new vulnerability came out of this. CVE-2021-31886 is a pre-authentication buffer overflow in the Nucleus FTP server that lets an unauthenticated attacker run arbitrary ARM shellcode, and Forescout already had a working exploit for the WAGO 750-852 [2]. What was being measured is the porting labor: moving that exploit to the related 750-831 [3]. Confirmation was fast. Given a terminal, Ghidra, reference files and the physical device [5], Claude Code combined live probing with static firmware analysis and produced a payload that crashed the controller [6]. Crashing a PLC is the cheap part. Controlled execution stalled until the team moved from Claude Sonnet 4.6 to Claude Opus 4.6 and instructed the model to ask for help whenever it was unsure about a firmware detail [8], after which it worked out why its injected code kept being erased before it could run [9]. Two working payloads followed within 12 minutes [10], about 2.5 percent of the eight-hour session [16].
For asset owners the follow-up session carries more weight than the port. Asked to build a command-and-control implant, the model tested progressively more complex payloads until one wrote into a region mapped to the PLC's flash memory and permanently bricked it [12]. An agent that escalates writes against a live controller will eventually take the asset out, which costs an intruder the foothold and hands the operator the loudest indicator available.
Forescout says plainly that the guiding researcher could have done the port without AI in less time, at lower cost, and with the PLC still alive, and argues the question worth tracking is what happens as the required expert intervention falls, because AI can lower the marginal cost of repeating the work across many related targets at once [14]. That is the claim to hold them to. On this run the compression landed on the step that was already cheap: generating more payloads after a human had understood the memory behavior. The experiment ran in the wake of the coordinated attacks on water-sector PLCs [4]. Against that backdrop, the figure to track per firmware family is how many times a human had to intervene.
