NSA, CISA, FBI Name Six Chinese AI Firms in Systematic Model Distillation Advisory
Every frontier lab's API access control is now a national security lever that the labs themselves cannot enforce. OpenAI and Anthropic have no mechanism to distinguish a Moonshot researcher using a proxy from a legitimate foreign customer; the advisory describes a problem (systematic extraction) that rate-limiting or usage caps cannot solve, because the Chinese firms are already spreading requests across accounts, platforms, and third-party aggregators. The realistic countermeasure is either blanket geographic restriction (which breaks legitimate business in allied nations and the EU) or Commerce treating API consumption at scale as a controlled export—a rule that has never existed and would require rewriting the EAR to govern outputs, not inputs.
OpenAI, Anthropic, and Google now face a choice: enforce terms-of-service bans on military and state-actor use (which this advisory shows are being routinely circumvented), or accept that API-layer access is the binding constraint on Chinese AI capability development—and that constraint is failing.
The distillation playbook identified here (DeepSeek extracted five ChatGPT versions and four Claude versions to train R1 and R3; Moonshot extracted 18 U.S. models including Anthropic's Fable 5) means the chip embargo's core assumption—that fab scarcity blocks military AI—is already obsolete in the field. Watch whether Commerce moves to API-layer controls (treating high-volume token consumption itself as a controlled export) or leaves enforcement to the labs' terms of service, which this advisory proves are unenforceable against state-scale adversaries using proxies and gray-market resellers.
Did the White House direct the advisory's release now, or is this NSA/CISA/FBI moving independently to create a record for future enforcement action?
Fault Lines
Strategic intelligence, synthesized daily — with a public track record. Every call graded against what actually happened.