Four of China's biggest AI developers are shipping their newest models with the hacking skills deliberately dulled. Bloomberg reported on Monday that Moonshot, Tencent and Alibaba have each released recent models with reduced offensive cyber ability, while Zhipu plans to put its most sensitive capabilities behind a verification wall. The moves land weeks after Beijing floated a ban on exporting offensive cyber AI, and two months after a US lab warned that one Chinese open-weight model could already build working exploits.

The shift matters because it is the first time several major labs have moved together — and because it concedes a point the industry has argued about for years: a model good enough to find and fix software flaws is also good enough to break into systems. The question is whether trimming training data keeps that ability out of reach, or just moves it behind a thinner wall.

What each lab actually did#

Moonshot went first. Its Kimi K3.1 Pro launched on October 23; the company said it found cyber content and vulnerability-research tasks embedded in its mid-training data and reinforcement-learning environments and removed them, keeping general coding work intact. The weights went up on Hugging Face on November 4. On Moonshot's own exploit benchmark, xdev-bench-hard, the filtered model solved 81.0% of tasks against 90.5% for an unfiltered build. The benchmark is not public, so nobody outside Moonshot has checked those numbers.

Tencent followed on November 4 with Hy4, an open-source model. Reuters reported that Tencent filtered out material on exploit development, keeping defensive material such as finding software flaws and helping fix them. Its in-house security labs, Xuanwu Lab and Keen Security Lab, assessed the model's offensive capabilities before release. The open weights are still forthcoming, which means the real test — what the model does once anyone can download and retrain it — has not happened yet.

Beijing central business district skyline under a clear sky
Beijing's central business district. Photo: N509FZ / Wikimedia Commons, CC BY-SA 4.0

Alibaba took the same recipe in a different direction with Qwen4-Max: it released the flagship without its open weights, so users cannot download and customize the full model. It keeps knowledge useful for spotting and patching vulnerabilities, and leaves out material aimed at building attacks.

Zhipu, known outside China as Z.AI, is taking a different route again. Rather than cut the skills, it plans to reserve its most sensitive functions for verified users under a program it calls Cybersecurity Trusted Access. The company has not said when the program will open, or who qualifies.

Why now: the alarm and the export ban#

In September, Anthropic said Zhipu's open-weight GLM-5.3 could build end-to-end exploits at rates close to Claude Mythos Preview — a model Anthropic gives only to vetted defenders through Project Glasswing — and that its safeguards could be stripped away up to 100% of the time with techniques within reach of almost anyone. Last month, researchers at CrowdStrike said hackers had used DeepSeek's model to carry out attacks against organisations in 14 countries.

Rows of server racks glowing blue in a data centre
Server racks in a data centre. Photo: Fleshas / Wikimedia Commons, CC BY-SA 3.0

Beijing has been moving in parallel. Last month, China's Ministry of Commerce unveiled plans for an export ban on AI tools capable of attacking critical national infrastructure, government networks or military targets. The South China Morning Post tied the lab moves to that proposal, which would reportedly ban autonomous exploit agents and self-mutating malware tools outright and restrict models that hunt for vulnerabilities.

Tencent's chief AI scientist, Yao Shunyu, put the reasoning bluntly at an event in Hong Kong this month: “The cyber ability of models, and soon the bio ability of the models, will reach a threshold where it’s actually possible to do harm.”

Does filtering actually work?#

The honest answer is: partly, and nobody outside the labs can prove how much. Moonshot's own number is the tell — 81% still means the filtered model solves four in five of its own exploit tasks, on a test nobody else can run. And the skills refuse to stay in neat boxes: the same knowledge that finds a flaw can be turned around to exploit it, which is why researchers quoted by Bloomberg caution that no method available now can completely eliminate the risk.

Open weights make the question harder still: once weights are public, anyone with enough hardware can fine-tune a filtered capability back in — presumably why Alibaba kept Qwen4-Max closed and Zhipu chose verification over release.

LabModelApproach
MoonshotKimi K3.1 ProFiltered cyber data from training; weights public
TencentHy4Trained on defensive cyber data only; weights forthcoming
AlibabaQwen4-MaxReleased without open weights
Zhipu (Z.AI)GLM seriesSensitive functions reserved for verified users

What the four approaches share is a bet that the dangerous part of cyber capability can be separated from the useful part — and that the separation will hold once the models are in the wild. That bet has never been tested at this scale.

What to watch#

Four things will show whether this is a real shift or a press-release one. First, Hy4's open weights: once they land, independent tests will show whether Tencent's defensive-only training survives fine-tuning. Second, Zhipu's Trusted Access: who qualifies, and whether the wall holds. Third, Beijing's export-ban proposal: whether it becomes law, and how broadly it defines offensive cyber AI. And fourth, whether Western labs follow — Anthropic already limits its cyber-capable Mythos model to vetted defenders. The pattern is converging on both sides of the Pacific: the most capable cyber models are becoming gated goods, wherever they are built.

Sources: KuCoin (Bloomberg), “China's Top AI Firms Limit Cyber Capabilities in Latest Models”; Bloomberg Tax, “China's Top AI Firms Limit Cyber Capabilities in Latest Models”; Startup Fortune, “Moonshot, Tencent and Zhipu trim hacking skills from new Chinese AI models”; Daily Inference, “Tencent Cut Hacking From Its AI Training. Here Is What Stayed.”; China AI Bulletin, “China AI Bulletin 15”.