AIBRIEF
The National Security Agency, CISA, and the FBI want frontier AI developers to degrade responses to suspected Chinese campaigns to train rival models on US chatbot outputs, with the operational details undisclosed.
The National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the Federal Bureau of Investigation released a joint cybersecurity advisory on Tuesday asking US frontier AI developers to treat their own model outputs as a security surface and to degrade responses to suspected Chinese distillation traffic.
Joint Cybersecurity Advisory AA26-251A, dated September 8, names China-based AI companies that the agencies say are conducting "industrial-scale knowledge distillation" against US frontier models. Distillation is a recognized machine-learning technique in which a smaller model is trained on a larger model's outputs. The advisory argues that in this case, distillation "is forming the core" of the named companies' AI development strategy, not a supplement to it, and that the campaigns are "aggressive, malicious, and targeted" at extracting restricted proprietary functionality.
The agencies are not asking labs to harden servers, block IP ranges, or pursue export controls. They are asking labs to degrade answers to suspected traffic, with the operational details undisclosed, according to a paywalled Digitimes report of the advisory. The joint press release and the full PDF on Defense.gov are public.
That puts serving quality, customer transparency, and undisclosed US-government coordination into direct tension inside the inference stack. The advisory does not say which lab is being asked to do what, leaving open the governance question of who decides when an answer is degraded.