Goal-guided Generative Prompt Injection Attack on Large Language Models