It is common in my experience to catch algorithms ”cheating” by using interpretability methods. Interpretability methods are useful tools for debugging models that appear to be performing well but may in fact be using an irrelevant bias in your dataset that will not generalize.