Anthropic's September 2026 Threat Intelligence Report details systematic "model distillation" attacks by Chinese AI firms, with Alibaba's Qwen accessing Claude 151 million times across 3,500+ accounts in May–July 2026 to extract internal reasoning processes (Chain of Thought). Attackers used prompt injection exploits to bypass safeguards; the report flags links to Chinese military/security agencies and reveals similar campaigns by Moonshot AI (Kimi) and three unnamed entities. This IP theft accelerates China's AI capability parity while exposing a critical vulnerability in API-based model protection—actionable for any SME relying on or building LLM services.