JULI: Jailbreak Large Language Models by Self-Introspection