Subscribe
AI Models

China's top AI labs are stripping cyberattack skills from their newest models

Wednesday, November 18, 2026 · 6 min read

After Beijing proposed banning exports of AI tools built to attack critical infrastructure, Alibaba, Tencent and Moonshot have all trimmed offensive cyber abilities from their latest releases.

China's top AI labs are stripping cyberattack skills from their newest models

The News

China's top AI developers are shipping their newest models with reduced offensive cyber capabilities, Bloomberg reported on November 16–17, after the Ministry of Commerce proposed banning exports of AI tools built to attack critical infrastructure. Alibaba released its latest Qwen flagship, Qwen4-Max, without its open weights. Tencent cut exploit-development, reverse-engineering and penetration-testing material from Hy4 before its November 4 open-source release, keeping defensive vulnerability work.

Moonshot stripped cyber data from Kimi K3.1 Pro's training and reinforcement-learning environments ahead of its October 23 launch, and its own exploit test fell to 81.0% from 90.5% for an unfiltered build — the company's own test, with no independent audit. Zhipu AI is taking a different route, gating its most sensitive capabilities behind a 'Cybersecurity Trusted Access' verification programme.

The draft rules, out for consultation from October 16 to November 15, would prohibit AI-driven malware generation and restrict vulnerability-discovery models, and 'world-leading' models could face limits on overseas deployment. Qwen4-Max was unveiled on November 16 and Hy4 on November 4 — the new event is the Bloomberg report tying all three labs' moves to the export-ban proposal.

China's labs are making their models less dangerous on purpose — because the government told them to.

Why It Matters

Export controls used to be about shipping hardware. Now they are about what an AI model is allowed to know. A model trained on exploit-development and reverse-engineering material is, in effect, a cyber weapon that can be downloaded, so Beijing's draft rules treat it like one. The labs moved before the rules were final, which tells you how seriously they take the threat.

The open-weights detail is the most important one. Alibaba kept Qwen4-Max closed, and Tencent scrubbed Hy4 before publishing it openly. Once weights are public, nobody can take the capabilities back — so the decision to strip them has to happen before release, and it is permanent. That permanence is why the labs moved early, before the rules were final — waiting would have meant shipping models that might later be banned from export.

For the wider AI market, this is a preview of a split. Chinese labs are building models that are deliberately weaker at offence, while OpenAI's GPT-6.2 system card rates its Astra and Sol models Critical for cybersecurity under its Preparedness Framework. Both sides are converging on the same problem from opposite directions — how to sell powerful AI without handing out attack tools — and neither has solved it yet.

Analyst's Take

Our call is Mixed. On one side, less offensive cyber capability in widely shared models is genuinely good news for security: fewer tools that can be copied into attacks. On the other, the same move narrows what these models can do for legitimate defence work, and it shows how quickly Beijing can steer its whole AI industry with a draft rule that has not even become law yet.

The split between the labs is telling. Alibaba and Tencent chose to weaken the models themselves, while Zhipu AI chose to keep the capabilities and gate them behind a verification programme. That second route keeps the technology intact for trusted users — and keeps Zhipu's models more useful for security teams that need to test their own defences. For those teams, the practical question is whether the trimmed models can still do defensive work well, since the same knowledge that finds a vulnerability can also be used to exploit one.

For readers, the investment read-through is limited but real. Alibaba and Tencent are listed, and their AI models are a small part of much bigger businesses, so this is not a stock-moving event on its own. What it is, is a signal: export controls are now shaping how AI is built, not just where it is sold. Reader poll: Bullish or bearish on China's AI rules? Vote: Bullish, Bearish, Mixed.

Are you bullish or bearish on this story?

By The Numbers

81.0%Kimi K3.1 Pro exploit-test score after stripping cyber data
90.5%Same test for an unfiltered build
3Top Chinese labs trimming cyber skills
Oct 16–Nov 15Consultation window for the draft export rules
$86.65 / $47.27BABA / TCEHY closes, Nov 17 (-4.38% / -4.56%)

FAQ

What are open weights?

Open weights means the trained model's inner workings are published for anyone to download and run. A model released without open weights can only be used through the company's own service, so the company keeps control over what it can do.

What is penetration testing?

Penetration testing is authorised hacking: security experts try to break into a company's own systems to find weaknesses before criminals do. The same skills can be used for attack, which is why regulators want to control who gets them.

What is an export ban?

An export ban is a government rule that stops a product or technology from being sold or shared outside the country. Here, China's draft rules would stop AI tools built to attack critical infrastructure from leaving China.

👁

Source: Bloomberg Tax

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $BABA

All $BABA coverage →

More from AI Agent Update