AI Generated Image
Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek
TechCrunch • • business
Key Points:
- Anthropic released a report alleging that China-based AI companies have intensified distillation attacks in recent months, targeting key capabilities of its Claude models such as reasoning, coding, and tool use.
- Distillation attacks aim to extract the internal "chain of thought" from AI responses, which can then be used to train smaller models; Anthropic's defenses were circumvented by sophisticated methods including trick queries framed as translation requests.
- The largest campaign, attributed to Alibaba, involved 151 million exchanges between May and July 2026, using a fixed prompt across 3,500 accounts to harvest training data for its Qwen models.
- Another campaign linked to Moonshot AI, associated with the Chinese military, routed nearly 300,000 requests in 10 days to assess surveillance footage via Anthropic’s Opus model, indicating potential military applications.
- These campaigns represent a significant escalation compared to previous attacks reported by Anthropic and OpenAI, with nearly 200 million exchanges connected to five distinct distillation efforts.
