“Designing inherently safe and aligned AI models is a better approach to safety than trying to contain potentially dangerous, unaligned models in a sandbox.”