Best practices when running a benchmark on online models [D]

I'm developing a benchmark for a low resource language and I don't want it to be leaked and used for training when it is being used to get predictions. For locally run models it shouldn't be a problem, but for models that are only accessible via API, it is.