A Single Neuron Is Sufficient to Bypass Safety Alignment in LLMs