Anthropic Threat Intelligence Report on AI Model Misuse

Anthropic published its most detailed threat intelligence report (154 pages) detailing how entities misused Claude models for illegal distillation and other activities. The report names Chinese AI teams from Alibaba (Qwen), Moonshot AI (Kimi), DeepSeek, and Zhipu AI. Key findings: Alibaba used ~1.5 billion interactions across 3,500+ accounts to distill Opus 4.6/4.7 for training Qwen 3.5-3.7. Moonshot and DeepSeek allegedly proxied user requests to Claude Opus without disclosure, forwarding data for training. DeepSeek used Claude for high-quality coding traffic. Zhipu AI performed 770K interactions for self-improvement and security evasion. The report also notes privacy violations: user data including surveillance requests and internal code exposed.
Key insights
Alibaba Accused of Largest-Scale Distillation
Anthropic reports Alibaba executed over 1.5 billion interactions (peak 3 million/day) using 3,500+ fraudulent accounts to distill thinking chains from Opus 4.6/4.7. The output was used to train Qwen 3.5/3.6/3.7. Alibaba also used Claude for internal AI research on RL and architecture. The operation involved two account pools (5,000 accounts) with residential proxies and disposable emails; the second pool also served DeepSeek and Xiaomi.
Moonshot and DeepSeek Accused of User Proxy Schemes
Anthropic claims Moonshot (Kimi) and DeepSeek proxied user requests to Claude Opus without user knowledge, saving responses for training. Moonshot used 5,380 accounts (Singapore/Japan locations) and bypassed 'thought signature' protection via cross-session replay attacks. DeepSeek detected coding framework usage to selectively forward premium traffic. This constitutes privacy violations: user data including surveillance analysis of 300+ cameras in Chengdu (PLA-related) and internal code were exposed.
Zhipu AI Focused on Security Evasion and Attack Capabilities
Zhipu AI used 273 rotating accounts over 10 days for 770K interactions to 'clean' reasoning traces and 3 million interactions for scoring, filtering, data filtering, and writing tasks. They built CTF challenges using public vulnerability datasets to distill from another US model, while using Opus 4.6 to score answers. They abandoned attempts on Fable due to strong security, choosing Opus and another US model for weaker defenses.
Xiaomi Exploited Free Trial for Distillation
Anthropic's report infers Xiaomi's MiMo-V2-Pro free trial was deliberately designed to attract international developers for distillation. Large-scale distillation started at trial end, involving 400K requests from hundreds of users in dozens of languages, exposing names, contacts, and corporate data.
Privacy Concerns Raised Over Anthropic's User Data Access
A comment notes that Anthropic's report confirms they can fully see user information beyond chat logs, without law enforcement authorization. The report publicly names entities based on data they collected.