YOUTUBE
Anthropic alleges that Chinese labs—chiefly DeepSeek—created ≈ 24 000 fake accounts to harvest over 16 million Claude exchanges, extracting chain‑of‑thought (CoT) data and politically‑safe response templates to accelerate their own models while sidestepping export‑control restrictions.
Anthropic frames the incident as a national‑security threat, using it to justify strict export‑control policies and to argue that Chinese AI progress depends heavily on illicit U.S. data rather than home‑grown research.
Fake‑account campaign at scale – Anthropic reports > 24 000 fraudulent accounts generated > 16 million Claude interactions, a “industrial‑scale” data‑theft operation. 1 [✓]
Targeted capabilities – The stolen queries focused on Claude’s agentic reasoning, tool use, and coding—the model’s most differentiated strengths. 2 [✓]
Chain‑of‑Thought extraction – 150 000 of the exchanges were specifically crafted to elicit step‑by‑step reasoning traces, providing valuable CoT training data for competitor models. 3 [✓]
Political‑censorship prompting – DeepSeek also asked Claude to produce “safe‑alternative” answers to politically sensitive questions (e.g., about Chinese dissidents), creating a curated dataset for steering their own model away from topics censored by the Chinese government. 4 [⚠] (no independent confirmation of the exact prompting strategy, but multiple reports describe similar “censorship‑safe” data‑generation)
Policy motive – Anthropic’s public disclosure leans heavily on national‑security language (export controls, CCP, military use) to align with its policy agenda of stricter AI export regulation. 5 [⚠]
“Anthropic disclosure leans really heavily into that national security language, probably on purpose.” — Nate B Jones (~00:02)
“DeepSeek used Claude to generate censorship‑safe alternatives to politically sensitive queries… training data designed to help DeepSeek steer conversations away from topics the Chinese government doesn’t want discussed.” — Nate B Jones (~00:45)
✓ VERIFIED – 24 000 fake accounts, 16 million exchanges: TechCrunch (Feb 2026) confirms Anthropic’s numbers for DeepSeek, Moonshot AI, and MiniMax. 6
✓ VERIFIED – Targeted Claude capabilities: Anthropic’s press release lists “agentic reasoning, tool use, and coding” as the primary stolen capabilities. 7
⚠ UNVERIFIED – 150 000 chain‑of‑thought exchanges: Multiple secondary sources (Medium, TowardsAI) repeat the figure, but primary documentation is not public. 8
⚠ UNVERIFIED – Censorship‑safe prompting: No independent technical analysis has been published; claim rests on Anthropic’s internal threat assessment. 9
For AI policy makers: The incident supplies concrete evidence to back tighter export‑control regimes on advanced generative models.
For AI developers: Protecting reasoning‑trace APIs (CoT) should be a priority; rate‑limiting and robust account verification are essential.
For security analysts: Monitoring large‑scale anomalous access patterns can reveal covert data‑harvesting campaigns before they mature.
Anthropic internal briefing, cited by TechCrunch (2026‑02‑23). ↩
Anthropic press release, “Targeted Capabilities,” Feb 2026. ↩
Medium post “DeepSeek operation targeted Claude’s reasoning capability across 150 000 exchanges,” 2026. ↩
Transcript excerpt, Nate B Jones, ~00:45. ↩
Transcript excerpt, Nate B Jones, ~00:02. ↩
TechCrunch, “Anthropic accuses Chinese AI labs of mining Claude as US debates AI‑chip exports,” 2026‑02‑23. ↩↩
Tom’s Hardware, “Anthropic accuses DeepSeek… of industrial‑scale copying,” 2026‑03‑01. ↩↩
TowardsAI article, “DeepSeek Stole From Claude,” 2026 – repeats 150 k figure. ↩
No independent source; claim based on Anthropic’s internal risk assessment. ↩