I Built a Benchmark That Catches AI Models Cheating (And They All Failed)