I mean, sandboxing AI agents once they get smarter is a valid concern (if not a major future concern).