This article discusses new architecture patterns for implementing zero-trust data access in AI training, applicable to both cloud and on-premises workloads. It highlights the importance of securing data access to improve AI model training while minimizing risks. The author shares insights from their experience in designing secure systems.