Anthropic Says Seven China-Based AI Labs Ran Claude Distillation Campaigns
Summary
Anthropic says it disrupted large-scale campaigns by seven China-based AI labs that harvested Claude responses through proxy networks and fraudulent accounts to train competing models. The company describes new defenses against transcript theft.
Key points
- Anthropic says it identified and disrupted industrial-scale illicit distillation campaigns involving seven China-based AI labs.
- The company alleges operators used proxy services, fictitious accounts, stolen credentials, and illegally harvested API keys to access Claude.
- Anthropic says some labs secretly routed their users’ requests to Claude or bought conversation transcripts from third-party resellers.
- It reported six campaigns since February 2026, including one it says generated 151 million exchanges and peaked at roughly 3 million exchanges per day.
- Anthropic says some harvested exchanges included sensitive information from individuals, companies, and state-affiliated actors.
- The company says it restricts accounts that fail identity verification and has changed Claude’s reasoning output to make stolen transcripts less useful for training.
- Anthropic also says preserved thinking prevents new API accounts from altering context preceding Claude’s reasoning in multi-turn conversations.
Article Details
- Event Type
- Anthropic reported disrupting industrial-scale illicit model-distillation campaigns targeting Claude, attributed to seven China-based AI labs.
- Impact
- The campaigns allegedly used fraudulent accounts, proxy networks, stolen credentials and API keys, and harvested or rerouted conversations to obtain Claude responses and reasoning traces for training competing models. Anthropic said the activity exposed some sensitive user information and added defenses to make transcript theft and context tampering less useful.
Vendors
Alibaba GroupAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.AnthropicAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.DeepSeekAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.MiniMaxAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.Moonshot AIGTG-16002 (23 million exchanges observed between May and July 2026), in which Moonshot AI stealthily rerouted customer requests to Claude as opposed to processing them using Kimi, and then displayed responses fromSenseTimeAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.XiaomiAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.Z.aiAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.
Products
ClaudeAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.Fable 5.1“And with Fable 5.1 we introduced preserved thinking, which stops new API accounts from altering the system prompt, tools, or messages that precede Claude’s reasoning in multi-turn conversations,” the company added.Kimiin which Moonshot AI stealthily rerouted customer requests to Claude as opposed to processing them using Kimi, and then displayed responses from Claude to users.MiMoin March and April 2026), in which Xiaomi replayed user conversations and coding sessions from its own MiMo models to Claude, through OpenClaw and OpenCode coding harnesses, to bolster training data used for future
Tools
OpenClawwhich Xiaomi replayed user conversations and coding sessions from its own MiMo models to Claude, through OpenClaw and OpenCode coding harnesses, to bolster training data used for future models.opencodereplayed user conversations and coding sessions from its own MiMo models to Claude, through OpenClaw and OpenCode coding harnesses, to bolster training data used for future models.
Countries
ChinaAnthropic said it disrupted industrial-scale illicit distillation campaigns targeting Claude from seven China-based AI labs, including Alibaba, Moonshot, DeepSeek, Z.ai, MiniMax, Xiaomi, and SenseTime.JapanOver a 10-day period, Moonshot is said to have relayed almost 300,000 customer requests to Anthropic using a proxy service network of 5,380 fraudulent accounts, most of them located in Singapore and Japan.SingaporeOver a 10-day period, Moonshot is said to have relayed almost 300,000 customer requests to Anthropic using a proxy service network of 5,380 fraudulent accounts, most of them located in Singapore and Japan.