The secret redirection of user queries across thousands of fraudulent accounts to extract the capabilities of Claude models represents a flagrant violation of usage policies and intellectual property rights.
Article index:
- Anthropic’s investigative report and accusation details
- Fraud mechanism and the network of fake accounts in Asia
- Knowledge distillation operations to train the Kimi chatbot
- DeepSeek and Xiaomi involvement in similar practices
- Joint warning by US intelligence and security agencies
- Official denial from Beijing and warnings of countermeasures
- Frequently asked questions
Anthropic’s investigative report and accusation details
Leading American artificial intelligence research and model company “Anthropic” published a detailed and strongly worded report on Thursday, accusing Chinese artificial intelligence startup “Moonshot AI” of secretly routing nearly 300,000 requests and queries from its users to “Claude” models—specifically the advanced “Opus” model—without the knowledge or consent of those users.
The American company confirmed that Moonshot exploited the advanced software and cognitive responses produced by Claude to train and improve its chatbot known as “Kimi,” in what Anthropic described as a “systematic and carefully calculated knowledge distillation campaign to steal advanced software capabilities without incurring exorbitant development and research costs.”
Fraud mechanism and the network of fake accounts in Asia
The new report adds precise evidence and details to accusations that Anthropic first raised last February, when it revealed that Chinese laboratories had created more than 16 million interactions with Claude models across approximately 24,000 fraudulent and coordinated accounts aimed at extracting weights and knowledge.
According to Anthropic’s forensic investigations, Moonshot used an advanced network comprising 5,380 fake accounts, whose IP addresses and geographic locations were disguised to appear as though they were operating from Singapore and Japan to bypass the strict geoblocking policy imposed by Anthropic to prevent access to its services and models from inside Chinese territory.
Knowledge distillation operations to train the Kimi chatbot
Anthropic observed an organized and extremely rapid behavioral pattern that raises suspicion, as Moonshot redirected extraction traffic to any new model released by Anthropic in less than 24 hours of its official launch, confirming that it possesses an operational infrastructure prepared and set up in advance to track American updates rather than being merely random, individual attempts.
This technical practice is known as “knowledge distillation,” where a developer harvests and aggregates the outputs of a superior and advanced artificial intelligence model to train a weaker model and improve its responses at a negligible computational cost, allowing the Kimi chatbot to deliver performance that rivals Claude in code generation and solving complex logical problems without investing in training massive foundation models.
DeepSeek and Xiaomi involvement in similar practices
Anthropic’s accusations were not limited to Moonshot alone; the report also named two prominent Chinese companies, “DeepSeek” and computing and phone giant “Xiaomi,” confirming their involvement in distillation practices and unauthorized extraction of the same models through digital channels and intermediary accounts.
This data reveals a hidden battle raging for control of weights and algorithms, as rising Chinese companies rely on the response efficiency and logical reasoning of American models to compensate for strict technological restrictions imposed on the export of advanced chips and computing accelerators to China, transforming Claude models—against the will of their developers—into secret tutors for competing models.
Joint warning by US intelligence and security agencies
The release of the report coincided with an unprecedented official security move in Washington, as the Cybersecurity and Infrastructure Security Agency, in cooperation with the Federal Bureau of Investigation and the National Security Agency, issued a joint official warning accusing six Chinese companies of extracting “billions of tokens” from leading American systems.
American agencies described these practices as “aggressive and malicious distillation activities carried out on a large industrial scale and targeting secret technology transfer,” warning of their repercussions on national security and technological superiority ahead of an anticipated bilateral summit between the leaders of the two countries at the end of this month.
Official denial from Beijing and warnings of countermeasures
In contrast, Moonshot, DeepSeek, and Xiaomi declined to comment immediately on the report, while China’s Ministry of Commerce rushed to affirm that “there is no factual or legal basis” for the American accusations regarding unauthorized industrial distillation.
Beijing warned that it would take decisive countermeasures if Washington moves to “contain and suppress leading Chinese companies,” emphasizing that its country’s technical superiority rests on legitimate and independent domestic research efforts.
Frequently asked questions
Question: What was the direct accusation leveled by Anthropic against the Chinese company Moonshot?
Answer: It accused them of secretly passing 300,000 user requests to the Claude model to extract its responses and train its Kimi chatbot.
Question: How did the Chinese company manage to bypass the geoblocking imposed by Anthropic?
Answer: It used a complex network of 5,380 fake accounts registered through servers in Japan and Singapore to evade blocking restrictions.
Question: What is the concept of knowledge distillation in training artificial intelligence models?
Answer: It is the process of extracting outputs and answers from a super-intelligent model and using them to train a smaller model and improve its performance at a low cost.
