While large language models present real risks to society, experts say they can be tested and largely controlled using well-worn cybersecurity and policy choices.