How are they going to stop it?

A certain combination of nonstandard characters will make an AI character drop an n-word no problem

I guess they could chuck the output through whisper or something to see if it transcribes back to anything dodgy?

LLM security feels very ball of sand held together with duct tape haha