Stealthy Attack on Large Language Model based Recommendation