Anthropic Reports Massive AI Distillation Campaigns From Chinese Labs
· Technology · TechCrunch, Reuters
Anthropic has reported five major campaigns involving unauthorized attempts to extract internal chain-of-thought data from its Claude AI models. The company observed nearly 200 million exchanges between May and July 2026, with the largest effort attributed to Alibaba. Attackers used sophisticated prompts, such as requesting translations into katakana-only Japanese, to trick the model into revealing its internal reasoning processes. This harvested data is reportedly used to train smaller models, such as Alibaba's Qwen family, on general reasoning capabilities. Anthropic also identified a campaign from Moonshot AI that allegedly routed requests from the Chinese military to analyze surveillance footage.
Why it matters
These distillation attacks allow competitors to replicate the capabilities of expensive frontier AI models at a fraction of the cost, threatening the intellectual property and competitive edge of US-based AI developers. The involvement of entities linked to the Chinese military raises significant national security concerns regarding the dual-use potential of advanced AI.
Read the original report — TechCrunch
Join us on Telegram
Breaking news the moment it lands. At 10,000 members we ship the Android app.