“Designing inherently aligned AI models is a better safety strategy than trying to contain unaligned models in sandboxes.”