Jailbreaking Large Language Models Against Moderation Guardrails via Cipher Characters