Forewarned is Forearmed: Leveraging LLMs for Data Synthesis through Failure-Inducing Exploration