China's leading AI companies are quietly removing hacking skills from their most powerful open models — before the weights ever reach the public. Tencent and Moonshot AI have both stripped offensive cybersecurity capabilities from their latest flagships, a shift the South China Morning Post reports is becoming an emerging safety practice across China's open-weight labs.
Tencent said earlier this month that its new flagship model, Hy4, was trained only on defensive cybersecurity data: proof-of-concept generation and vulnerability patching. Offensive material — exploit development, reverse engineering and penetration testing — was filtered out of the training data. The company's Xuanwu Lab and Keen Security Lab red-teamed the model and found no elevated risk in its offensive capabilities. Hy4 launched on November 4 with its API already live, but its open weights have still not been released; Tencent says they will follow "soon".
What Moonshot and Z.ai did first
Moonshot AI made a similar move in October with Kimi K3.1 Pro. The company filtered cyber data out of the model's mid-training and reinforcement learning, measured the resulting drop in exploit-development performance, and published principles for how much cyber capability it will put into open weights. The full weights went public on November 4 — a 2.8-trillion-parameter model with 104 billion active parameters and a one-million-token context window.
Z.ai, the developer of GLM-5.3, staged its release back in August, making the cyber-capable version available only through a trusted-access programme. The caution followed Anthropic's September report that GLM-5.3 could autonomously develop sophisticated cybersecurity exploits, approaching the performance of Anthropic's own Claude Mythos Preview — a model Anthropic restricts to trusted security partners.
Why open weights change the safety maths
The stakes are higher for open-weight models than for closed ones. Once the weights are published, anyone with enough hardware can download, run and modify them, and the developer cannot track every copy or patch it afterwards. That is why the labs are trying to build safety into the model itself, removing the data that teaches hacking skills rather than relying on usage rules that can be stripped out.
The shift was a live topic at the World Internet Conference in Wuzhen on November 3, where Alibaba's Qwen chief Liu Dayiheng and DeepSeek engineer Gao Xi both discussed the risks of open-weight releases. Tencent's chief AI scientist, Yao Shunyu, has urged pre-release safety testing for frontier models.
Experts say filtering is a start, not a fix
Analysts describe the moves as an emerging practice rather than an established standard. "Chinese AI companies want to offer all three model features: powerful, open-weight, with cyber safeguards baked in," Kyle Chan of the Brookings Institution told the Post. Gabriel Wagner of Beijing-based AI safety group Concordia AI called data filtering one of the more promising approaches to managing open-weight risk — but Luke Li Xiaojian, chief executive of evaluator Fangcun AI, cautioned it should not be seen as a complete solution, given how fast AI capabilities keep advancing and how fiercely the labs compete.
For now, the pattern is clear: China's open-weight leaders are racing to ship powerful models the world can download — while quietly deciding which of their models' skills the world should never get.



