FBI and NSA Accuse Chinese AI Firms of ‘Industrial-Scale’ Model Distillation
The NSA, CISA and FBI have accused several Chinese artificial intelligence companies of conducting large-scale knowledge distillation campaigns against leading US AI models, including ChatGPT, Claude, Gemini and Grok.
According to the agencies, the activity has been taking place since at least 2024 and has involved millions of requests routed through different accounts, APIs, proxies, cloud services and third-party aggregators.
Agencies Call Distillation a Core Development Strategy
The US agencies said the campaigns were designed to extract proprietary capabilities from American AI systems and use them to improve Chinese models.
They described the activity as “industrial-scale” and said knowledge distillation was not simply being used as an additional development technique, but as a central part of some companies’ AI strategies.
Distillation generally involves using the outputs of a more capable model to generate data that can help train another system.
DeepSeek Allegedly Used Multiple US Models
DeepSeek was specifically accused of using frontier US models to generate synthetic training data for its R1 and R3 models.
According to the advisory, the company used outputs from four versions of Claude, two versions of Gemini, five versions of ChatGPT, and Grok 4.
Those models allegedly helped DeepSeek improve areas including agentic capabilities, question-and-answer performance, creative writing, and professional writing.
Moonshot AI Also Named
Moonshot AI was accused of distilling 18 different US models to help train its Kimi-K2 and Kimi K3 systems.
The models allegedly included Fable 5, described in the report as Anthropic’s most advanced commercially available model.
The company reportedly sent millions of queries aimed at extracting capabilities related to agentic reasoning, coding, data analysis, computer vision, broader logical reasoning, and visual processing.
Requests Allegedly Routed to Avoid Detection
The advisory also describes a network of techniques allegedly used to make the activity harder to detect.
These included distributing requests across multiple accounts, models and platforms, as well as using native APIs, remote cloud providers and third-party aggregators to obscure user metadata.
The companies were also accused of using proxies and gray-market technology services to bypass geographic restrictions, platform terms of use, and safeguards built into frontier AI systems.
US agencies said these techniques allowed large volumes of requests to be spread across different channels while reducing the likelihood that individual AI providers would identify the broader extraction campaign.
The post FBI and NSA Accuse Chinese AI Firms of ‘Industrial-Scale’ Model Distillation appeared first on ProPakistani.
