UNITED States government agencies have alleged that several Chinese AI firms conducted “industrial-scale” distillation campaigns to extract and repurpose proprietary capabilities from leading US frontier models. In a joint advisory published on 8 September 2026, the FBI, NSA and CISA claim that companies such as DeepSeek, Moonshot AI, MiniMax, StepFun and Z[.]AI used outputs from US models to train their own proprietary systems and to cut the costs and time required to develop new AI capabilities.
The advisory states that these firms reportedly accessed billions of tokens across millions of exchanges from models including variants of Claude, GPT, Gemini and Grok since at least late 2024.
The document describes a range of techniques used to scale these activities, including distillation methods that employ mature “teacher” models to train “student” models, along with industrial-scale practices such as chain-of-thought extraction, automated failover between pathways during attempts to block access, and sophisticated quality frameworks to detect countermeasures.
It also claims that the involved firms obtained bulk premium subscriptions for US AI models and routed distillation requests through native APIs, cloud providers and third-party aggregators to obscure metadata and evade safeguards, sometimes via transfer stations that circumvent geographic restrictions.
In terms of practical response, the advisory urges US AI companies to implement comprehensive detection and mitigation of anomalous behaviour, share telemetry with industry peers, and tune responses to suspected malicious attempts. Analysts emphasise continuous monitoring of API access patterns and collaboration with model providers to share indicators of compromise.