On the impact of contaminations in graphical Gaussian models