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.