Jailbreaking Large Language Models Against Moderation Guardrails via Cipher Characters Andy Zhou School of Information Sciences