Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates