The letter argues federal agencies should more rigorously test AI models for security threats and sandbox evasion, set new safety standards and lean into defensive AI.

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.