The aim of this paper is to address the validity of default probability models calibrated on a dataset including a very low (or none) number of defaults. The few approaches, proposed by the specialized literature are based on the confidence intervals computed via probabilistic, Bayesian or analytic methods. We propose a benchmark of these methods and we explore the cases for which they are the most suited for. We propose a new method for computing the upper bound of the default probability in low default portfolios, employing resampling approaches. For the special case of no default portfolios we investigate an adjustment based on non-Gaussian behavior of the distance to default. We address also the issue concerning the dependency between defaults using non-Gaussian copulas.