Large Language Models of Code Fail at Completing Code with Potential Bugs