Investigating Accidental Misalignment: Causal Effects of Fine-Tuning Data on Model Vulnerability — AI Alignment Forum