As of: 2026-10-13 07:38 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz. Canonical page: https://postcutoff.com/v/better-stack-deepseek-kimi-sent-prompts-to-claude/ # DeepSeek and Kimi Secretly Sent Your Prompts to Claude Better Stack, 13 September 2026, YouTube. 71,214 views as of 13 October 2026. Kind: Review. Watch: https://www.youtube.com/watch?v=hSVCYzphZjQ ## Why it is here 8-minute explainer of Anthropic's September 2026 threat intelligence report, which accuses seven Chinese labs including DeepSeek, Moonshot and Alibaba of distillation attacks on Claude (~71k views). The title's claim is the channel's reading of the report. ## Description (written by Gemini from the video) **Summary** A presenter on the Better Stack channel reviews the findings of Anthropic’s September 2026 threat intelligence report regarding widespread "illicit distillation" by major Chinese AI labs. The video details how companies like Alibaba, Moonshot AI, DeepSeek, Zhipu AI, Xiaomi, SenseTime, and MiniMax allegedly harvested Claude’s reasoning data and proxied live customer queries to Claude Opus, as well as the new technical safeguards introduced in models like Claude Fable 5.1. --- ### **What is shown** - **[00:00]** Intro showing logos of targeted labs (DeepSeek, Kimi, Qwen, Zhipu/GLM). - **[00:07]** Excerpts from Anthropic’s September 2026 threat intelligence report detailing unauthorized proxying and sensitive government/military use cases. - **[00:35]** Diagram explaining legitimate model distillation (teacher model, prompt set, training set, student model). - **[00:54]** Overview graphic of illicit distillation infrastructure (fake accounts, stolen credentials/APIs, residential proxies, and transfer stations). - **[01:21]** Diagrams of Alibaba’s attack: injected prompt wrappers extracting inline `` chain-of-thought traces from Claude Opus 4.6 and 4.7 into Qwen 3.5, 3.6, and 3.7. - **[02:41]** Moonshot AI (Kimi) section: architecture diagrams showing customer queries forwarded to Claude Opus and the "cross-session replay attack" that recovered raw reasoning from thinking signatures. - **[04:14]** DeepSeek section: diagrams showing how incoming API headers (`claude-code`, `claude-agent-sdk`, `opencode`) were detected to selectively route developer traffic to Claude Opus. - **[05:25]** Analysis of Zhipu AI (GLM), Xiaomi (MiMo), SenseTime’s reseller data purchasing, and MiniMax’s proxy shell company. - **[07:06]** Anthropic's mitigation updates: CoT summarization, encrypted "preserved thinking" in Claude Fable 5.1, adversarial classifiers, and regional identity checks. --- ### **Claims & numbers** - **Report findings**: The presenter reviews Anthropic’s September 2026 threat intelligence report documenting covert, large-scale distillation. - **Prior findings**: In February 2026, Anthropic accused DeepSeek, Moonshot, and MiniMax of using over 24,000 fraudulent accounts across 16 million exchanges. - **Alibaba (Qwen)**: Allegedly ran the largest distillation attack measured, with 151 million exchanges (peaking near 3 million/day) across 3,500 fraudulent accounts between May and July 2026 to distill reasoning into Qwen 3.5, 3.6, and 3.7. - **Moonshot AI (Kimi)**: Forwarded ~300,000 customer requests to Claude Opus in a 10-day period using 5,380 fraudulent accounts, totaling over 23 million exchanges from May to July 2026; bypassed thinking-signature protections using cross-session replay attacks. - **Data leaks**: Relayed requests exposed sensitive data to Anthropic, including Russian Ministry of Defense database credentials, Chinese municipal Public Security Bureau surveillance matching tools, and Chengdu CCTV surveillance queries. - **DeepSeek**: Relayed ~12.1 million exchanges over 14 days in July 2026, selectively targeting users running coding harnesses (Claude Code, Claude Agent SDK, OpenCode) to harvest agentic coding data. - **Zhipu AI (GLM / Z.ai)**: Rotated 273 accounts for 3.4 million exchanges over 17 days; attempted to extract cyber capabilities from Claude Fable before abandoning it due to Fable’s cyber safeguards. - **Xiaomi (MiMo)**: Replayed users' coding sessions across ~400,000 requests and 1,500 accounts, reportedly launching and extending a MiMo-V2-Pro free trial to harvest international developer interactions. - **SenseTime & MiniMax**: SenseTime purchased scraped Claude transcripts from third-party proxy resellers; MiniMax operated a proxy shell company exclusively reselling US frontier models to harvest logs. - **Safeguards**: Anthropic implemented reasoning summarization, classifiers against extraction prompts, mandatory identity verification for accounts originating from China, Russia, or Iran, and "preserved thinking" in Claude Fable 5.1 to lock prompts and encrypt reasoning blocks. --- ### **Notable quotes** - **[00:00]** *"Anthropic just took shots at DeepSeek, Kimi, Qwen, GLM, and basically every Chinese model..."* - **[01:25]** *"Operators affiliated with Alibaba ran the largest distillation attack we have ever measured..."* - **[02:48]** *"Moonshot AI silently forwarded customer requests to Claude, instead of processing them using Kimi."* --- ### **Assessment** This is an explainer and news review breaking down the technical and geopolitical details of Anthropic's published threat intelligence report. The presenter relies entirely on Anthropic's released statistics, diagrams, and case summaries without performing hands-on reproduction of the alleged exploits. _Described by gemini-3.8-flash on 2026-10-13 from the video's audio and frames._ ## Related - 2026-09-10: [Anthropic report details AI-orchestrated cyberattacks and distillation by Chinese labs](https://postcutoff.com/e/2026-09-10-anthropic-threat-intelligence-report-sept-2026/)