@misc{jain2024what, title={What Makes and Breaks Safety Fine-tuning? A Mechanistic Study}, author={Samyak Jain and Ekdeep Singh Lubana and Kemal Oksuz and Tom Joy and Philip H. S. Torr and Amartya Sanyal and Puneet K. Dokania}, year={2024}, eprint={2407.10264}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2407.10264} }