Prompt as Triggers for Backdoor Attack: Examining the Vulnerability in Language Models