This article explores different sandboxing techniques for executing AI code safely. It discusses the limitations of containers, the advantages of gVisor and microVMs, and the importance of policy design to prevent data leaks. The author provides a decision-making framework to choose the right sandbox based on threat models and operational needs.