Skip to main content
See every side of every news story

Anthropic Says Alibaba, DeepSeek Used Claude to Train Rival Models

China

Martin LELIEVRE / AFP via Getty Images/Getty

Martin LELIEVRE / AFP via Getty Images/Getty

What Happened

Anthropic alleged China‑based AI firms, including Alibaba, Moonshot and DeepSeek, ran 'illicit distillation' campaigns routing Claude outputs into rival training pipelines. It said it disrupted the campaigns, banned accounts and attributed nearly 200 million exchanges, over 151 million linked to Alibaba between May and July 2026.

What Happened

Anthropic alleged China‑based AI firms, including Alibaba, Moonshot and DeepSeek, ran 'illicit distillation' campaigns routing Claude outputs into rival training pipelines. It said it disrupted the campaigns, banned accounts and attributed nearly 200 million exchanges, over 151 million linked to Alibaba between May and July 2026.

Where Sources Disagree

  • arrows_outputFraudulent Account Count Discrepancy: Some outlets report that DeepSeek, Moonshot, and MiniMax utilized 24,000 fraudulent accounts for distillation campaigns. Conversely, other reports cite a smaller network of 5,000 to 5,380 accounts used by Moonshot and DeepSeek.
  • arrows_outputMoonshot Distillation Volume: Reports vary on the total volume of distillation exchanges attributed to Moonshot AI, with some sources citing over 3.4 million exchanges while others report the figure reached over 23 million during the same period.
  • arrows_outputDistillation Attack Figures: Reports vary slightly regarding the total volume of distillation attacks identified, with some sources citing nearly 190 million exchanges while others report the figure closer to 200 million.

Timeline

September 11, 2026

September report and responses: On September 11, 2026 Anthropic released a Threat Intelligence report naming Alibaba, Moonshot, DeepSeek, Xiaomi and Zhipu and saying the five labs launched nearly 190 million distillation exchanges (part of ~200 million total across five campaigns); the company said it had banned associated accounts, added safeguards (including producing less-detailed reasoning transcripts) and engaged authorities. The disclosures prompted joint accusations by U.S. agencies and raised the prospect of sanctions or Entity List actions for firms found to have crossed into IP theft.

July 31, 2026

May–July large-scale campaigns: Between May and July 2026 Anthropic observed massive distillation activity: operators linked to Alibaba generated more than 151 million Claude exchanges (peaking near 3 million/day across ~3,500 fraudulent accounts), Moonshot accounted for over 23 million exchanges, and DeepSeek produced more than 12 million exchanges over a 14-day span. Anthropic said these campaigns targeted Claude's advanced capabilities and in some cases relayed live customer conversations without users' knowledge.

June 05, 2026

June letter to lawmakers: In a June letter to U.S. senators Anthropic's head of policy said Alibaba illicitly ran roughly 28.8 million exchanges with Claude between April 22 and June 5, calling for legislation to curb such attacks. That disclosure preceded the wider threat-intelligence report and signaled the issue moving into policy debates.

Summary by Ground AI

Sources

See All 58

A varied selection of sources chosen by Ground to reflect the diversity of this story’s coverage.

View All Sources

Similar News Topics

News
Feed Dots Icon
For You
Search Icon
Search
Blindspot LogoBlindspotLocal