Enhancing LLM Agent Safety via Causal Influence Prompting