They can use a subset of the training data as a validation set. Given enough attempts at the test set it's very easy to overfit your model to that test set, me

They can use a subset of the training data as a validation set. Given enough attempts at the test set it's very easy to overfit your model to that test set, meaning that although the accuracy looks higher, it would generalize worse on a different subset of the data.