Anthropic accuses DeepSeek, Moonshot and MiniMax of ‘distillation attacks’
16M+ Claude exchanges via ~24,000 fraudulent accounts
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 4 October 2026
Your AI and this story
- GPT-6.2 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 4 HydrogenIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- Per lab (Anthropic): DeepSeek over 150,000 exchanges, Moonshot AI over 3.4 million, MiniMax over 13 million; total over 16 million via ~24,000 fraudulent accounts
- Method: ‘hydra cluster’ networks of proxy services and fraudulent accounts to get around Anthropic’s ban on access from China; attribution ‘with high confidence’ through IP address correlation, request metadata and infrastructure indicators
- Targets: agentic reasoning, tool use and coding, which Anthropic calls Claude’s most differentiated capabilities
- MiniMax’s campaign was caught while active; when a new Claude model launched, MiniMax ‘pivoted within 24 hours, redirecting nearly half their traffic’ to it
- Response: classifiers and behavioral fingerprinting for distillation patterns (incl. chain-of-thought elicitation), stricter verification for education, security-research and startup accounts, sharing indicators with other labs, cloud providers and authorities, and model-level safeguards; ‘no company can solve this alone’
- Follow-ups: Alibaba’s Qwen lab accused of a larger campaign (28.8M exchanges, June 2026); the Sept 2026 threat report named seven China-based labs
What happened
Anthropic published a security post that described coordinated campaigns by three Chinese labs to query Claude at scale through fake accounts and proxy resellers, harvest its outputs and use them as training data. It named the labs, gave per-lab exchange counts and listed its countermeasures.
Why it matters
“Distillation attacks” became a named security and policy issue in 2026. The report framed copying a frontier model’s outputs as theft of capabilities rather than ordinary competition, and later US moves (the April OSTP distillation memo, the Hagerty–Kim sanctions proposal) built on that framing.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 1 primary, 1 press, 1 reaction
Primary
- Anthropic: Detecting and preventing distillation attacks (Feb 23, 2026)anthropic.com, official
Press
- Infosecurity Magazine: Chinese AI firms hit Claude with distillation attacks, Anthropic warnsinfosecurity-magazine.com, press
Reactions
- Hacker News discussionnews.ycombinator.com, discussion
Changes
- Filed (pre-cutoff background)