On the (Non-)Robustness of Two-Layer Neural Networks in Different Learning Regimes