“A mitigation for emergent misalignment during fine-tuning is to provide the model with a benign explanation for the task, which prevents it from adopting a generally malevolent persona.”