Even a human support tech shouldn't have that much power. That should be a request to a completely different team that gets hard off of telling people "No".

It's not impossible. It's actually pretty simple. You just don't give it access to credentials or resources or sensitive information of any kind. Does that make them basically useless? Yep. But then you don't create massive security risks and legal liabilities with gross negligence either, so... gotta weigh those things against each other.

noone seems to know how to properly secure them, probably because it is impossible

Most likely because all the AI engineers are idiots without any coding experience to begin with.