Towards Understanding the Fragility of Multilingual LLMs against Fine-Tuning Attacks