A New Front in the AI Wars: Chinese Companies' 'Distillation' Operation Against Claude Code

The global AI race is becoming more complex than ever with hardware embargoes between countries and cyber intelligence wars. In particular, the chip wars between the US and China are pushing Chinese tech giants toward unusual and controversial methods to stay in the innovation race.

Cornered on the hardware side by the strict export restrictions the US administration has imposed on high-performance AI chips (especially Nvidia's flagship GPUs), Chinese companies are resorting to "shortcuts" to close the quality gap with their rivals. The latest and most striking example of this came to light with a massive infiltration attempt targeting Anthropic's flagship model.

24,000 Fake Accounts, 16 Million Conversations

Leading American AI research company Anthropic publicly disclosed organized data theft against its systems on an unprecedented scale. According to the statement, DeepSeek, Moonshot AI and MiniMax, the hottest Chinese AI platforms of recent times, carried out an extensive "data raid" on Claude Code, Anthropic's advanced coding and analysis assistant.

Working in an organized way to get past the system's defenses, these platforms created a total of 24,000 fake accounts on Claude Code. Managing this network of fake accounts with bots and automation tools, the Chinese companies generated a full 16 million different conversation scenarios on the system. The goal was to collect Claude Code's high-quality responses and use this massive dataset to train their own open-source and local AI models.

In this way, the Chinese platforms copied, at almost no cost, the coding, reasoning and problem-solving capabilities their US rivals had reached by spending hundreds of millions of dollars. As a result, they were able to offer users in the market much cheaper yet high-performing subscription plans.

A Dangerous Weapon: "Distillation"

At the center of this operation is a technique known in the AI literature as "distillation" (model distillation). The distillation technique aims to train a smaller, weaker "student" model using the high-quality outputs produced by a large, high-capacity, highly advanced "teacher" language model (here, Claude Code).

What makes the incident truly dangerous has less to do with costs and more to do with safety protocols. When developing their models, Western companies build deep safety barriers (guardrails) against military use, biological weapon recipes or destructive technologies. However, rival models trained through "distillation" can take only the raw technical knowledge and capability while easily bypassing the moral and ethical restrictions of the original model. Experts warn that, as a result, the leaked capabilities could be made directly available to Chinese military organizations or cyber operations units.

Anthropic's Countermove Was Comprehensive

After detecting the data theft, Anthropic officials immediately took down the network of fake accounts on the system. But the company did not limit its defense to that:

1. Rapid Intelligence Sharing: The source of the leak, the behavior patterns of the accounts (IP blocks, query patterns) and the bot techniques used were immediately shared with other trusted, whitelisted AI developers (OpenAI, Google) and cloud storage providers. 2. Notifying Official Channels: Because the situation carried national security risks, the relevant government authorities were brought into the process. 3. Security Architecture Updated (Claude Code Security): Anthropic announced that it had integrated new defense mechanisms into its platform that will not disrupt the ordinary user experience but will block excessive data extraction (scraping) and distillation techniques.

This latest scandal once again lays bare the risks of opening models, open or closed, to the outside world, and clearly shows that in the years ahead AI security will perhaps be the hottest front in cyber defense.