The fact that secret behaviors can be well concealed within a ML generated model makes it critical to start defining what can and can't be tried.