Anthropic reports large-scale AI distillation campaigns
Anthropic says China-based AI companies have conducted increasingly sophisticated attempts to extract capabilities from its Claude models. In a report released Thursday, the company said it identified five campaigns involving nearly 200 million exchanges. The targeted capabilities included agentic behavior and tool use, coding, data analysis, and logical reasoning. OpenAI has previously described similar activity and linked it specifically to DeepSeek.
Anthropic said distillation efforts seek to obtain a model’s chain of thought and use it to train a smaller system through supervised fine-tuning. Although Claude normally presents users with summarized thinking rather than internal reasoning traces, attackers reportedly found ways to prompt the model to reveal them. One method disguised the request as a translation task involving previous working memory and katakana-only Japanese.
The largest campaign was attributed to Alibaba. Anthropic recorded 151 million exchanges from May through July 2026, involving 3,500 accounts and reaching almost three million exchanges a day. Another campaign, linked to Moonshot AI, routed nearly 300,000 requests through 5,000 accounts over ten days, mainly targeting Claude Opus. One request sought analysis of surveillance footage for abnormal behavior and appeared to originate through a Chinese military network.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.