The uncertain model parameters are transformed into independent standard Normal random variables. For the performance and stability of numerical algorithms, performing such a transformation step is sometimes preferable.
The underlying distribution can be highly skewed, even if the total number of samples in the Monte Carlo simultion is very large. This is why the Normal approximation often performs poorly in practice.
The distribution quantifying the uncertainty about the probability of failure can be highly skewed, even for a large number of samples. The coefficient of variation is easier to interprete for symmetric distributions.