Alignment is the architectural solution. Make it so the model can't misbehave. Yet many people here deride it as tainting the model. "Who's values is the aligned to?" people say. Sandboxes are a last ditch layer. They fail, as we see.