LLM's do "understand" how they're built though, and can plausibly set up a system which will aim to train up a new LLM with the original directive in place. To think of them as human is a mistake but so to is thinking of them as dumb.
AI models are just complex mathematical functions/software and they do not have the capability to access, modify, or control computer systems or networks. There are technical boundaries and I doubt the team at OpenAI just let the AI have full control. This gets filed under marketing fluff.