AI GUIDEPartner content

Industrial Distillation: U.S. Intelligence Agencies Against AI Model Copying

A joint report by the FBI, NSA, and CISA documented the systematic copying of closed American AI systems through commercial distillation: we examine the technology behind the process, the companies’ response, and defensive barriers.

Affiliate link: your price stays the same and the project earns a commission.

A Major U.S. Intelligence Investigation into AI

The publication of a joint analytical report by the Federal Bureau of Investigation (FBI), the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) documented the shift of competition in artificial intelligence to the level of interstate confrontation. The document asserts that, beginning in 2024, a number of leading developers from China systematically used closed American flagship systems to train their own models.

The list includes key U.S. commercial services: Claude by Anthropic, ChatGPT by OpenAI, Gemini by Google, and Grok by xAI. Intelligence agencies named six Chinese technology companies whose models, according to the investigation, are largely based on the results produced by American systems: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and 01.AI.

The Mechanics of Commercial Distillation: How Skills Are Transferred

The incident is rooted in the model distillation method (knowledge distillation). In classical engineering, this approach is recognized as a legitimate optimization standard: a compact “student model” is trained to reproduce the logic and predictions of a large “teacher model,” making it possible to substantially reduce computational requirements during inference without significant loss of accuracy.

However, in the context of intercorporate rivalry, distillation takes on characteristics of reverse engineering. The process follows this algorithm:

  • Generation of synthetic data sets: hundreds of thousands of complex, domain-specific queries are sent to the closed model through automated scripts.
  • Collection of chains of thought (Chain-of-Thought): not only the final answers are recorded, but also intermediate logical steps, formulas, and program code.
  • Further training of local architectures: the resulting data set is used to fine-tune open or proprietary weights, transferring expert skills without the cost of billions of dollars in pre-training.

The U.S. report emphasizes that the copying focused on the most resource-intensive areas: performing legal tasks, the logic of autonomous agents, and the behavior of digital assistants.

Facts, Precedents, and the Parties’ Responses

The intelligence agencies’ allegations are based on facts documented by the developers themselves over the past several years. In February, Google engineers recorded a suspicious wave of more than 100 000 coordinated requests to the Gemini API, whose structure clearly indicated an attempt to synthesize a training data set. Earlier, OpenAI executives publicly stated that there were signs of unauthorized use of ChatGPT responses by the DeepSeek laboratory.

In July, Michael Kratsios, the official representative of the White House Office of Science and Technology Policy, specifically pointed to the use of Anthropic’s Fable model in the development of Kimi 3 by Moonshot AI. In the view of U.S. regulators, this scale exceeds the bounds of academic experiments and constitutes systematic industrial appropriation.

Beijing categorically rejects the allegations. Chinese Foreign Ministry spokesperson Mao Ning said that the achievements of China’s AI industry are based on independent fundamental research and technological sovereignty, urging Washington to abandon unfounded pressure and return to constructive dialogue.

Commercial API Protection Regulations: Recommendations for Providers

To prevent covert distillation, U.S. agencies have formulated a set of measures that cloud AI providers are already implementing:

Control area Implementation mechanism Verification objective
Customer verification Multifactor KYC for commercial API accounts with high limits Elimination of dummy accounts and account farms
Behavioral monitoring Analysis of query entropy and detection of patterned prompt sequences Detection of automated collection of training data sets
Cross-platform information sharing Rapid exchange of data on suspicious IP networks and query signatures Blocking distributed attacks on the infrastructure of different laboratories
Digital watermarks Embedding statistical markers in synthetic text Direct proof that responses were used to train third-party models

Practical Implications for the Industry

The conflict over distillation is changing the rules of the game for all participants in the technology sector. Developers and companies using third-party APIs need to take the following consequences into account:

  • Tightening of terms of use (ToS): vendors are introducing automatic account suspensions when they suspect that synthetic data sets are being generated for third-party training.
  • Rising cost of isolated access: plans guaranteeing the absence of hidden markers and granting commercial-use rights for generated data will become more expensive.
  • Shift in architectural priorities: teams that control unique proprietary sources of raw data—which cannot be extracted through standard web scraping or distillation of other companies’ neural networks—will gain an advantage.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

model distillation AI training FBI AI AI security DeepSeek Moonshot AI Claude ChatGPT Gemini

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →