It was just giving commands to run that were plain wrong and extremely destructive, and unless you already knew what they were doing you were screwed.
Here: https://chatgpt.com/share/696539b6-65f0-8010-9324-5e35da42ee...
I have 4-5 more conversations like this. It's honestly almost a piece of art, the LLM keeps spouting out shit like "Ah got it, your issue is clear now", and digging deeper into the wrong direction.
Email/DNS I outsource, everything else I homelab.
In the end I literally had to give up, it's just too problematic.