As long as you have a sufficient test suite you could probably run a Ralph Wiggum loop and have it brute force it. Creating the test suite would be harder though.
The phrase "sufficient test suite" is doing a LOT of work here. You would need to know what the data from every sensor is supposed to be along with how every piece of the machine is supposed to perform. AI isn't going to be able to iterate into those parameters over night.
I don't think the farmer that just wants to switch his dead light bulbs out for generic ones without waiting for john deer certified tech surely is very proficient in writing a full test suite.
I'm not anti right-to-repair by any means. I'm just saying that "we don't need the manufacturer, just vibe code your own firmware" is not a solution to this problem.
Farmers should be able to swap parts, tweak loads of things in software, etc without software locks. I should be able to do that on my car too. But the way to achieve that is not through spending a few hundreds of thousands of dollars on Opus [latest] and praying that doesn't kill me, that's the part I have an issue with in OPs comment.
I've never been fond of the argument that there should be a professional software engineer certification, but hearing people like you being presented with the potential dangers and just going 'oh yeah just go with a better test suite and you can just wing it' makes me seriously reconsider.
Vibe code administrative systems for your local golf club to your hearts desire for all I care, god forbid somebody will have to stand around a bit longer before going for their 9 holes. But safety critical equipment is not the place to fuck around with the code prediction machines that have existed for 4 years, have been writing more-or-less acceptable code for 2, and will still regularly refer to themselves as MechaHitler or just make up shit. "Yes you're absolutely correct, I was wrong" doesn't help you one bit if you have just been chewed up by heavy machinery, and the fact that people like you exist who go 'oh just a few more more unit tests surely will fix it' is a terrifying thought.
But don't humans make mistakes too? Like are we sure the failure rate of AI with the right checks and bounds is lower than humans, who are flawed machines themselves?
If you need assurances, have a different LLM write the test suite.
'passing this test is too complex, let me find another way to implement your feature successfully'. If you never had this message or similar on opus 4.5, and never saw the resulting code, you wouldn't understand why some software engineer don't trust AI with security.