Case anthropic-distillationModel & data leakUnited States
Anthropic accuses DeepSeek, Moonshot and MiniMax of distilling Claude via 24,000 accounts
Anthropic said three Chinese labs used fraudulent account networks and proxies to run 16M+ Claude exchanges to extract capabilities for their own models.
HighAttack on AIStatus: Under investigationVerified case file
Open in the interactive mapCite this case
- Reported
- Feb 23, 2026
- Location
- United States
- Target
- Anthropic Claude
- Severity
- High
- Triggered by
- DeepSeek, Moonshot AI and MiniMax (per Anthropic)
- Actor type
- Criminal
- Reported loss
- None reported
- CVE
- None
What happened
Anthropic said three Chinese labs used fraudulent account networks and proxies to run 16M+ Claude exchanges to extract capabilities for their own models.
Impact: MiniMax over 13 million exchanges, Moonshot over 3.4 million, DeepSeek over 150,000, targeting coding, agentic reasoning and tool use. No direct financial loss disclosed.
Who triggered it
Actor: DeepSeek, Moonshot AI and MiniMax (per Anthropic)
Attribution: Alleged by Anthropic based on account, infrastructure and request-pattern evidence; not independently adjudicated
Motive: Extracting Claude capabilities to train competing models cheaply (intellectual property theft)
How it happened
Attack vector: About 24,000 fraudulent accounts and proxy services used to run over 16 million Claude exchanges in violation of terms and regional restrictions
Root cause: API access allows output harvesting at scale; account verification and regional controls bypassed with fake accounts and proxies
Timeline
- Campaigns run; MiniMax activity detected while it was still training a model
- Anthropic publicly accuses the three labs and details per-lab exchange counts
Response
Anthropic banned the accounts, published findings and cited the case in support of chip export controls; the labs were asked for comment.
Current status: Under investigation
Sources (open in a new tab)
- The Decoder the-decoder.com
- Interesting Engineering interestingengineering.com
- eWeek eweek.com
Cite this case
Anthropic accuses DeepSeek, Moonshot and MiniMax of distilling Claude via 24,000 accounts. AI Threat Watch, case anthropic-distillation, Feb 23, 2026. https://aithreat.io/cases/anthropic-distillation/
Permanent link: https://aithreat.io/cases/anthropic-distillation/