Five distillation campaigns targeted Claude, Anthropic says
Anthropic published a report Thursday accusing China-based labs of harvesting Claude’s reasoning to train their own models. One campaign, attributed to Moonshot AI, the maker of Kimi, appeared to route requests coming from the Chinese military.
Across five campaigns Anthropic counted nearly 200 million exchanges. The largest, attributed to Alibaba, ran 151 million exchanges from May to July 2026, producing training material for the Qwen models.
The Moonshot campaign sent nearly 300,000 requests through 5,000 accounts over one 10-day period, mostly against Opus. One asked Claude to review closed-circuit surveillance footage and judge whether the subject was “behaving abnormally”. OpenAI has reported similar activity and attributed it to DeepSeek.
Up next
Codex harness goes managed in the new Agents API
“Choosing a self-hosted sandbox does not make the Agents API ZDR-eligible,” OpenAI writes in the docs for its new managed Codex harness service, which currently supports data residency only in the United States.
Related stories
- OpenAI ties a reasoning-extraction campaign to Moonshot AI
- Irregular CEO blames human error for AI agent test hacks
- Anthropic’s FlashProve proves Llama 4 Maverick at 110 tok/s
- Tainted Claude Code build stayed in Cursor, Microsoft finds
- Cathedral’s agents broke into Anthropic in under 12 hours
- Claude gave up its sandbox binary and process memory
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.