Benchmarking AI Models by Their Ability for Subterfuge