PolicyTech
NSA, CISA and FBI name six Chinese AI firms for "industrial-scale" distillation of US frontier models
The Register / Engadget / Time News · 2026-09-09 · Source ↗
NSA, CISA and FBI name six Chinese AI firms for "industrial-scale" distillation of US frontier models
The Register / Engadget / Time News · 2026-09-09 · Source ↗
The National Security Agency, Cybersecurity and Infrastructure Security Agency and Federal Bureau of Investigation published a joint cybersecurity advisory Wednesday morning naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as running "aggressive, malicious, and targeted" knowledge-distillation campaigns against US frontier AI companies. It is the first US-government document to name Chinese AI companies collectively as intelligence targets rather than simply flagging one-off Terms-of-Service violations, and the advisory concludes that distillation "functions as the critical core" of the six companies' development programs rather than an ancillary training method.
The advisory covers billions of tokens exchanged across millions of API sessions since at least late 2024, spanning variants of Anthropic's Claude, OpenAI's GPT, Google's Gemini and xAI's Grok. It says the campaigns "likely operate with awareness of, and in some cases direction from" Chinese government entities. For DeepSeek specifically, the agencies claim the company targeted "reasoning capabilities, agentic functions, and specialized optimizations" from GPT-4, GPT-5 and multiple Claude versions to train its R1 and V3 models, and that DeepSeek's widely-cited $5.6 million R1 training cost is "misleading because it excludes the cost of data acquired through distillation." For Alibaba, the advisory claims "industrial-scale distillation" improved the Qwen family across all three of its 2025-2026 release cycles.
"The scale, coordination and persistence of the campaigns we describe is not consistent with commercial exploration. It is consistent with a national program to close a capability gap." — Rob Joyce, NSA Director of Cybersecurity, in the advisory's cover statement
The recommendation nobody expected
The document's second half contains an operational recommendation aimed at OpenAI, Anthropic, Google and xAI: US frontier providers should "quietly degrade" suspect Chinese-linked accounts rather than banning them outright, so as to inject noisy, subtly-wrong training signal into any downstream distillation. The recommendation is unprecedented in a joint cybersecurity advisory. Simon Willison flagged Wednesday evening that the recommendation may violate US providers' own published Terms of Service, since a "quiet degrade" policy would mean the provider knowingly furnishes lower-quality output to a paying customer without notification. Anthropic and OpenAI declined to comment on whether they intend to adopt the recommendation. Google's public response through Kent Walker acknowledged the advisory but committed only to "reviewing the pattern of usage the agencies describe."
Alibaba pushed back Wednesday afternoon. Eddie Wu, Alibaba Cloud CEO, told the FT that Qwen's training-data pipeline is "public, documented and reviewed by external auditors" and that any distillation from US frontier APIs would represent less than 0.4 percent of Qwen 3.8's total 12 trillion-token training set. DeepSeek did not respond. Moonshot AI, whose Kimi K3 flagship trained on 47,000 distillation trajectories according to an August paper the company published itself, said the paper's methodology was "misrepresented" by the US advisory and that the trajectories were sourced "from publicly available benchmark evaluation traces, not from paid API access." Z.AI's founder Tang Jie posted on Weibo that the advisory was "a political document, not a technical one."
What providers can actually see
The advisory's technical appendix lists the fingerprints US providers can use to identify likely distillation traffic: unusually-long context windows, extremely high sampling temperatures, output-token-heavy request patterns, VPN egress from Southeast Asian jurisdictions, and repeated resubmission of near-identical prompts to trigger sample diversity. Providers were also instructed to look for "coordinated prompt-family patterns" indicating an automated agent farm rather than human users. The document was co-published with the UK's National Cyber Security Centre and Japan's Cybersecurity Council, which is the first three-way NSA-NCSC-Japan alignment on a China-specific AI action.