You Can't Steal Nothing: Mitigating Prompt Leakages in LLMs via System Vectors