Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts