This paper proposes a novel, training-free defense method for LVLMs that amplifies their inherent safety capabilities by identifying and utilizing a single safe attention head to detect unsafe inputs and guide safer responses.