you could use llama guard or openai's omni-moderation model to flag bad actors