Modern AI applications deliver effective results only if the training data are of outstanding quality. Such data is a precious asset for both companies and public bodies, but this implies extraordinary efforts in protecting them from unauthorised access and reuse. In this workshop we will present some cutting-edge AI research from IDSIA in critical areas such as health and material sciences which involve the use of protected and restricted data. We will then explain how the CSCS Alps infrastructure, together with Apertus, can enforce data sovereignty in training AI models, also presenting a joint work from Artificialy and the Canton of Tessin in the automated translation of confidential material.