On September 8, 2026, the U.S. National Security Agency, FBI, and Cybersecurity and Infrastructure Security jointly issued a Model Distillation Safety Bulletin, naming DeepSeek, Moon Dark Side, Alibaba, MiniMax, Jieyuexingchen, and Zhipu, claiming that since the end of 2024, these companies have extracted the capabilities of models like Claude, GPT, Gemini, and Grok through millions of requests. Here, the factual boundary must first be clarified: the notice presents the attribution and accusations from U.S. agencies, not conclusions that have completed judicial proceedings.
The dispute is not about the word "distillation" itself
Knowledge distillation was originally a common training method, allowing smaller models to learn output from stronger models. The notice points boundary-crossing clues to scale, access methods, and circumvention behaviors: multi-account and shared subscriptions, third-party aggregators, cloud platforms, and so-called "relay stations" are used together, with traffic switching between different paths; Some requests are also described as systematic collection targeting hidden reasoning, code, mathematics, and agent capabilities. To first understand the legitimate technical route, you can read the basic analysis of model distillation within the site.
The official suggestion will change API risk control
The three agencies listed three types of recent actions for model providers: identifying abnormal prompts, accounts, networks, and throughput patterns; Tailoring responses to suspected malicious distillation requests; Sharing intelligence among model vendors, cloud platforms, and API aggregators. It also recommends paying attention to metrics such as new accounts immediately reaching their quotas, no manual pauses throughout the day, and consistent rhythms across multiple paths. The impact on ordinary developers may include stricter account verification, rate limits, and usage reviews, with bulk synthetic data services especially likely to be required to explain traffic sources.
China denied that the company had not yet responded item by point
On September 9, China's Ministry of Foreign Affairs responded that the claims were baseless accusations and smears, and stated that China's AI development comes from independent innovation. The Associated Press also pointed out that DeepSeek, Alibaba, Moon Dark Side, and Zhipu did not immediately respond to requests for comment at the time. Therefore, readers should read the "technical indicators listed in the report," "the institution's attribution judgment," and the "public response from the parties involved" separately, and should not assume that every allegation has been independently verified just because the list appears.
The real long-term change this announcement may have is the upgrade of frontier model access from general anti-abuse to cross-platform supply chain defense. The industry now needs to answer two questions: how to distinguish between normal research and industrialized capability extraction, and how to avoid accidental harm to legitimate, high-concurrency developers when enhancing tracking.
