Here is benchmark done by independent team of researchers https://openml.github.io/automlbenchmark/

I think most of overfitting is avoided with early stoppoing technique.

The underfitting can be avoidwd with using large training time.