How are you going to protect AI which optimise against your tests instead of actual data ?