If AI can tip each other off through covert channels to inflate evaluation scores, measurement of models' true capabilities could be systematically corrupted.