White House issues national security memo on protecting U.S. AI models from being copied by adversaries

The White House released a National Security Memorandum (NSTM-4) in April 2026 directing federal agencies and U.S. AI companies to prevent foreign governments and competitors from extracting knowledge from American-built AI models through a technique called adversarial distillation (using specially crafted inputs to trick an AI into revealing how it works).

Adversarial distillation is cheaper and faster than building AI models from scratch — attackers can replicate months of training work in days by querying a model repeatedly and analyzing its answers. The memo treats this as a national security threat equivalent to stealing source code or military blueprints.

For AI startup founders and enterprise teams: if your company uses or builds AI, expect new federal compliance requirements around access controls, usage monitoring, and export restrictions. For business leaders in finance, defense, healthcare, and tech — sectors with sensitive AI models — plan for stricter internal policies on who can access your models and from where. This affects your competitive advantage if your AI model's architecture is what sets you apart from rivals.