Uncertainty is Fragile: Manipulating Uncertainty in Large Language Models