When a model accidentally learns from information it shouldn't have access to, like future data or test set details.