This Is Your Doge, If It Please You: Exploring Deception and Robustness in Mixture of LLMs