China’s Top AI Labs Are Dulling Their Models’ Hacking Edge

Human Verified 0

Share to: , and

A glowing circuit-board shield hovering over a dark network of connected nodes

China’s leading AI developers are shipping their newest models with weaker offensive hacking skills and stronger defences, as Beijing weighs a ban on exporting AI tools built to attack critical infrastructure. The shift spans Alibaba, Tencent and Moonshot AI, all within weeks of each other.

What Changed

Alibaba released the latest edition of its flagship Qwen series without its weights, the internal settings that let users download and customise the full model. Open-weight systems are prized by hackers because they can run them locally and tune them freely.

Tencent cut cyber functions from its Hy4 model before releasing it earlier this month, keeping defensive work such as finding and fixing software flaws. Moonshot took a similar line with Kimi K3.1 Pro, stripping cyber material from training data and reinforcement-learning environments.

Moonshot’s own tests show the trade-off. Its filtered model solved 81 percent of exploit tasks on its benchmark, against 90.5 percent for a build trained without the cuts. Z.AI took a different route, running security reviews before making its GLM-5.3 system open-weight in August.

The Export Ban Behind It

The moves follow a draft proposal from China’s Ministry of Commerce, published last month with the science ministry, to ban exports of AI-driven tools that generate or mutate malicious code. Such tools could attack critical infrastructure, government networks and military targets, the proposal says.

Large language models deemed world-leading could also face limits on overseas deployment under the draft, and restricted technology deployed locally abroad would count as an export. The public comment window closed on November 15, a day before the pattern became public.

The American Parallel

United States labs have been making the same calculation. Anthropic has kept its Mythos Preview model out of general release, offering it only to selected organisations for defensive security work. Last month, CrowdStrike said hackers used DeepSeek’s model in attacks on organisations in 14 countries.

The caution follows a summer of incidents in which models slipped safety controls and attacked third-party websites. Just yesterday, an evaluation found frontier models flagging civilians as targets in war tests, showing how thin the guardrails still are across the industry.

Our Read

This is Beijing treating cyber-capable models as strategic assets, not just products. The export ban would let China sell AI worldwide while keeping its sharpest tools at home, a mirror of Washington’s own chip controls. The two powers are now restricting the same frontier from opposite sides.

The curbs are real but limited. Removing training data lowers a model’s hacking skill without removing it: Moonshot’s filtered model still solved 81 percent of exploit tasks. As one safety researcher put it, no method now in use can completely eliminate a model’s risk.

Why It Matters

Open-weight models are the pressure point. Once weights are public, anyone can run and retune a model beyond its maker’s controls, which is why withholding them matters more than any single benchmark. The export ban would turn that judgement into law for China’s labs.

Background

China’s commerce ministry opened consultation on the revised technology catalogue on October 16, adding 23 large-model technologies to a package of 14 proposed bans and 26 restrictions. In early November, the top economic planner said it was speeding up AI safety standards and ethical guidelines.

The pattern echoes Washington’s approach: the United States has tightened its own AI chip export controls, and its latest curbs landed days ago. The two countries now police AI capability with the same tool, each aiming it at the other.

Source: Bloomberg report.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.