LLMs? No. But LLMs are too slow for content moderation at scale.
Custom trained models? Maybe. Is the unicode characters in the training data?