Benchmarks for AI safety should test for undesirable properties like deception; no one should build systems capable of deception because such systems could circumvent other safeguards.