Did someone invent working LLM-based moderation? Serious question; it'd be interesting.
It actually does a better job than the stock "this comment does not go against our community standards" response you get from the human moderators of any social network.
Not so easy. Jailbreaks are becoming harder to perform every day.
Yes there are LLMs useful for such things and you could use them to make moderation decisions. YMMV with how "good" you want your moderation to be.