From Assistants to Adversaries: Exploring the Security Risks of Mobile LLM Agents