Multitask Mayhem: Unveiling and Mitigating Safety Gaps in LLMs Fine-tuning