US administration guardrails prevent the use of frontier AI models for certain tasks, including code security