I wonder how effective it would be to finetune a model to remove jailbreaks from prompts, and then use that as part of the pipeline into whatever agent