SignalSafety practiceSG-0070
Alibaba, Tencent and Moonshot cut offensive hacking data from new models as China weighs AI export curbs
Tencent filtered exploit development, reverse engineering and penetration testing out of the training data for its open-source Hy4 model. Moonshot removed cyber material from Kimi K3.1 Pro and says its score on its own unpublished exploit test fell from 90.5% to 81%. Alibaba kept defensive data but cut attack-focused data from Qwen4-Max and did not release its weights. The moves follow an October draft from China's Commerce Ministry to ban exports of AI tools able to attack critical infrastructure or government networks.
Why it matters
Open-weight Chinese models are attractive to attackers because they can be run locally and configured by the user, and CrowdStrike said in October that hackers used DeepSeek's model against organizations in 14 countries. Filtering training data raises the cost of misuse without removing it: open weights can be fine-tuned to try to restore skills, Moonshot's filtered model still solves four in five of its own exploit tasks, and Moonshot's figures have not been checked outside the lab.
Sources
Read the reporting
4 sources. Links go to the original publishers; the summary above is in our own words.
- newsChina’s top AI firms limit cyber capabilities in latest modelsThe Star (Bloomberg) · Nov. 17, 2026thestar.com.my/tech/tech-news/2026/11/17/chinas-top-ai-firms-limit-cyber-capabi…
- newsAlibaba launches Qwen4-Max flagship AI model without open weightsThe Next Web · Nov. 16, 2026thenextweb.com/news/alibaba-launches-qwen4-max-flagship-ai-model-without-open-w…
- newsMoonshot, Tencent and Zhipu trim hacking skills from new Chinese AI modelsStartup Fortune · Nov. 17, 2026startupfortune.com/moonshot-tencent-and-zhipu-trim-hacking-skills-from-new-chin…
- newsTencent rolls out AI model trained with limited cybersecurity data amid hacking fearsThe Star (Reuters) · Nov. 4, 2026thestar.com.my/tech/tech-news/2026/11/04/tencent-rolls-out-ai-model-trained-wit…
Signals
More from the news desk
Anthropic cut its disclosure deadline for AI-found high and critical bugs from 90 days to 30
OpenAI hardened GPT-6 Astra and Sol against jailbreaks by training on attacks its own model wrote
Anthropic found Claude could write harmless-looking training data that slips past its monitors and misaligns models
Cite and share
Use this signal
Citation
Paperclip Index. “Alibaba, Tencent and Moonshot cut offensive hacking data from new models as China weighs AI export curbs.” Signal SG-0070. Reported Nov. 16, 2026; updated Nov. 18, 2026. https://paperclipindex.com/signal/SG-0070