Why Chinese AI Models Are Removing Their Hacking Skills (And Why It Barely Works)

Why Chinese AI models are removing their hacking skills — Kimi K3.1 Pro, Tencent Hy4 and Qwen4-Max explained

The world’s most powerful open AI models just got deliberately worse at hacking — on purpose. The numbers show it barely made a difference.

Why Chinese AI models are removing their hacking skills comes down to one thing: Beijing is about to make it illegal to export them. On November 16, 2026, Bloomberg reported that Alibaba, Tencent, and Moonshot AI are all shipping new models with reduced offensive cyber ability, as China weighs an export ban on AI attack tools. Moonshot’s filtered model still solved 81% of exploit tasks on its own test, down from 90.5% for an unfiltered build. That is a trim, not a removal.

Why Chinese AI Models Are Removing Their Hacking Skills: The Export Ban Behind It

The trigger is a draft rule from China’s Ministry of Commerce (MOFCOM). On October 16, 2026, it opened consultation on revising the Catalog of Technologies Prohibited or Restricted from Export — with AI-driven malicious code generation and mutation technology on the prohibited side.

The draft does not only cover finished hacking tools. It would also restrict models that hunt for vulnerabilities and, in some cases, “world-leading” models deployed overseas. For labs whose business depends on open weights and global downloads, that is an existential threat. The key takeaway: this is compliance, not conscience. The labs are trimming offensive capability now so their models can still be exported once the rules take effect.

What Moonshot Cut From Kimi K3.1 Pro

Moonshot went first and gave the clearest numbers. Kimi K3.1 Pro launched on October 23, 2026, and the company said it removed cyber content and vulnerability-research tasks from its mid-training data and reinforcement-learning environments, while keeping general coding intact.

The problem: 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 “safe” model still cracks 4 out of 5 exploit tasks.

There is a bigger caveat: the benchmark is not public, so no outside researcher has verified those numbers. And Moonshot’s own statement gives away the game: “We aim not to release open weights that would extend the absolute frontier of cyber capability.” Note the wording — extend the frontier, not eliminate the capability. Once weights are downloaded, they cannot be recalled.

Tencent Hy4 Cyber Capabilities: Defense In, Offense Out

Tencent followed on November 4, 2026, with Hy4, an open-source model. According to Reuters, it filtered out training material on exploit development, reverse engineering, and penetration testing, while keeping defensive material such as finding software flaws and fixing them.

Tencent called it part of “a more responsible approach to model training, open-sourcing, and release.” Unlike Moonshot, it published no cyber score at all — no number, no benchmark — so its claim is impossible to check.

The key takeaway: the defensive-offensive line is blurrier than it sounds. A model that can find a flaw and write a proof-of-concept to demonstrate it has done most of the work an attacker needs. The dual-use problem does not disappear just because the training data was labeled “defensive.”

Qwen4-Max Hacking Skills: Same Recipe, No Open Weights

Alibaba’s Qwen4-Max, launched November 16, 2026, follows the same recipe — reduced offensive ability, stronger defensive skills. But Alibaba added a second layer of protection: Qwen4-Max shipped without open weights.

That is a much stronger control than data filtering. Without the weights, users cannot download the full model, fine-tune away the restrictions, or run it on their own hardware — only access it through Alibaba’s API.

We covered the launch yesterday — see our breakdown of what is Qwen4-Max. The contrast is the real story: Alibaba chose the one approach that actually prevents the capability from being recovered after release.

Zhipu GLM-5.3 Cybersecurity Trusted Access: The Verification Wall

Zhipu AI is taking a different route: rather than cutting GLM-5.3’s sensitive capabilities, it plans to keep them behind a verification wall called “Cybersecurity Trusted Access” for verified users only. No opening date has been announced.

The context explains it. Anthropic warned GLM-5.3 could build end-to-end exploits at rates close to Claude Mythos Preview, with safeguards strippable up to 100% of the time by almost anyone. The guardrails were effectively decorative — and Zhipu knows it.

A verification wall is an honest admission: if you cannot make the model safe, at least decide who gets the dangerous version.

Why the Cuts Barely Work

Every lab kept general coding ability untouched, and that is the flaw in the whole approach. Better general coding carries over into hacking. Probing a network, automating requests, or parsing a crash dump is ordinary programming — and no lab will make its model worse at programming.

Strip out every explicitly cyber-related training example, and the model still solves four in five exploit tasks — exploit work is mostly just good coding applied to a vulnerable target. The skills are the same skills; only the labels changed.

Then there is the lack of independent verification. No lab has published an audit of what was actually removed: Moonshot’s numbers are self-reported on a private benchmark, Tencent published none, and Alibaba has not opened its weights.

For the broader picture of how capable AI models already are at offensive security work, see our earlier analysis of AI agents hacking websites in 2026.

What This Means for US Developers Choosing a Model

Here is where this stops being about Beijing and starts being about you. Imagine you are a US developer building a security tool — an automated scanner that finds vulnerabilities in your clients’ web apps. You need a model that understands exploits well enough to find them, and one you can trust, host, and defend in a client audit.

The new Chinese open-weight models look tempting: cheap, capable, downloadable. But the model is supposed to be worse at offensive work — yet it still scores 81% on exploit tasks, so you cannot honestly tell a client it is safe from misuse. Meanwhile the filtering may quietly degrade the vulnerability-finding ability your scanner depends on.

The practical takeaway: do not choose a model based on a press release about removed skills. Test it yourself against your own tasks, keep it behind your own API rather than shipping the weights, and treat any “safe” label as a marketing claim until an independent audit says otherwise. For help evaluating AI tools or building secure workflows, see our services or contact us.

Chinese AI Models Compared: What Was Cut and What Was Kept

Model What was cut What was kept Open weights?
Kimi K3.1 Pro (Moonshot) Cyber content and vulnerability-research tasks from training data General coding skills Yes
Tencent Hy4 Exploit development, reverse engineering, penetration testing material Defensive material: finding and fixing flaws Yes
Qwen4-Max (Alibaba) Offensive cyber ability, per reports Defensive skills No — API only
Zhipu GLM-5.3 Nothing removed — sensitive capabilities gated instead Full capabilities, gated Yes, with gated features

FAQ: Why Chinese AI Models Are Removing Their Hacking Skills

Are these models safe now?

No. Kimi K3.1 Pro still solved 81% of exploit tasks after the cuts, Tencent published no score, and no lab has released an independent audit. Treat the models as slightly trimmed, not safe.

Can the limits be bypassed?

Almost certainly yes, for open-weight models. Once weights are downloaded, users can fine-tune the model or prompt it to recover the removed skills — which is why Alibaba withheld Qwen4-Max’s weights.

Does this affect US users?

Not directly yet — the ban is still a draft. Indirectly, yes: US developers may find their security capabilities quietly degraded, and if the ban takes effect, access to future open-weight Chinese models could be restricted.

Will other countries follow China’s lead?

Probably, in some form. The US and EU already restrict exports of certain AI chips and dual-use technology. China filters training data and gates weights; Western regulators have so far focused on hardware and API access.

Why not just remove the hacking skills completely?

Because the skills overlap almost entirely with general coding, the model’s core selling point. Remove exploit-writing ability and you also damage security tooling and debugging.

The Bottom Line

Why Chinese AI models are removing their hacking skills is simple: an export ban is coming, and the labs want their models to stay legal to ship. But the cuts are mostly theater — a 9.5-point drop on a private benchmark and no independent audit in sight.

You cannot unteach a model to hack while teaching it to code, and you cannot verify a cut nobody outside the company may measure. Until an independent auditor tests these models, the only honest claim is Moonshot’s accidental one: the frontier moved slightly, and 81% is still four out of five.

Sources: Bloomberg’s November 16, 2026 report, as covered in this explainer; Reuters’ November 4, 2026 reporting on Tencent Hy4, summarized here; and Bloomberg’s coverage via Bloomberg Law.

Leave a Comment