We propose a principled framework Bayesian Data Scheduler (BDS), which is an adaptive tuning-stage defense strategy against harmful fine-tuning with no need for attack simulation.