LLMs remain vulnerable to "jailbreaking" through adversarial prompts