Prompt Injection Attacks and Defenses in LLM-Integrated Applications