The September 8 joint advisory from the U.S. National Security Agency, FBI and CISA should be read carefully. Model distillation is a legitimate and widely used machine-learning technique. The dispute is over whether companies are using it to bypass controls and systematically extract restricted proprietary capabilities from frontier models.
That distinction matters because an overbroad political response would be as damaging as underreaction. Governments should not treat every mainland Chinese AI company, every researcher or every use of distillation as evidence of wrongdoing. But neither should they ignore credible, technically specific evidence of coordinated extraction when security agencies publish it.
The right response is a governance framework built around verifiable conduct: stronger API and account-abuse detection, rate and identity controls, model-output monitoring, cross-provider threat sharing, auditable enforcement standards and public technical indicators where disclosure is safe.
The U.S. agencies also argue that extracted model capabilities can accelerate military and cyber capabilities. That claim raises a legitimate national-security concern, but it should remain tied to evidence. Assertions about Mainland Chinese Communist Government awareness or direction should be attributed to the agencies unless and until additional public evidence establishes more.
For democratic governments, the policy objective should be resilience rather than collective suspicion: protect frontier-model capabilities, publish enough technical detail for independent scrutiny, coordinate with allies and industry, and distinguish ordinary research from deliberate efforts to defeat restrictions.
FCM’s view is that transparency strengthens national-security claims. When governments can release technical indicators, methodology and mitigation guidance without exposing sensitive intelligence, they should do so. That makes enforcement more credible and reduces the risk that strategic competition becomes a substitute for evidence.
Related coverage
Rapid Response: U.S. agencies warn of industrial-scale AI model distillation