Towards Safeguarding LLM Fine-tuning APIs against Cipher Attacks