Alibaba—Alibaba ran largest recorded unauthorized AI model extraction campaign against Anthropic's Claude to train Qwen
Anthropic and U.S. national security agencies (NSA, FBI, CISA) reported that Alibaba conducted a systematic 'distillation' campaign extracting Claude's reasoning transcripts and internal model-development data to train its Qwen models. Between May and July 2026, Alibaba-linked accounts made 151+ million unauthorized exchanges with Claude, peaking at nearly 3 million per day from 3,500+ fraudulent accounts using residential proxies, disposable emails, and virtual cards. Exchanges included sensitive information from individual users, multinational companies, and state-affiliated actors. On September 8, 2026, NSA/FBI/CISA issued a joint statement naming Alibaba (alongside DeepSeek, Moonshot AI, MiniMax, StepFun, and Z.AI) for systematically extracting proprietary AI knowledge from American firms at industrial scale since at least 2024.
Scoring Impact
| Topic | Direction | Relevance | Contribution |
|---|---|---|---|
| Intellectual Property Ethics | -against | primary | -1.00 |
| Overall incident score = | -1.288 | ||
Score = avg(topic contributions) × significance (critical ×2) × confidence (0.64)
Evidence (2 signals)
CNBC TV18: Anthropic report describes Alibaba's 'largest AI brain-theft operation ever recorded'
Anthropic's threat intelligence report detailed 151+ million unauthorized exchanges from 3,500+ fraudulent Alibaba-linked accounts between May and July 2026, extracting Claude's reasoning transcripts, RL environments, and model-architecture research to train Qwen 3.5/3.6/3.7.
Bloomberg: US says Alibaba, DeepSeek have 'systematically' siphoned AI models
NSA, FBI, and CISA issued a joint statement on September 8, 2026 naming Alibaba, DeepSeek, Moonshot AI, MiniMax, StepFun, and Z.AI for systematically extracting proprietary AI models from American firms at industrial scale via distillation since at least 2024.