Ai Engineering 7 min read

Chinese AI labs are trimming offensive cyber skills from new models as Beijing drafts export curbs

Bloomberg reported on November 16 that Alibaba, Tencent and Moonshot have shipped recent models with reduced offensive cyber capability after a Ministry of Commerce draft proposed banning exports of AI attack tools. The methods differ, and only Moonshot published a number.

China’s largest AI developers are releasing models with deliberately reduced offensive cybersecurity capability, Bloomberg reported on November 16 (also carried by Business Standard). The report names Alibaba, Tencent and Moonshot AI, and ties the pattern to a Ministry of Commerce proposal from October to ban exports of AI tools that can attack critical infrastructure, government networks or military targets.

The individual releases were already public. What the report adds is the framing: three labs made the same choice inside four weeks, under the same regulatory signal. For developers who self-host Chinese open-weight models, that affects what the next generation of weights can do and whether the largest models are released as weights at all.

Three labs, three methods

The labs did not do the same thing, and the differences matter for anyone evaluating the claims.

Moonshot (Kimi K3.1 Pro, October 23). Moonshot removed cyber-related data from mid-training and from its reinforcement-learning environments before the open-weight release, and it is the only one of the three to publish the cost. As we covered in our K3.1 Pro write-up, its in-house xdev-bench-hard score fell from 90.5% for a checkpoint without the cleaning to 81.0% for the shipping model. Moonshot said it does not intend to release open weights that push the offensive cyber frontier.

Tencent (Hy4, November 4). Tencent said Hy4 was trained with cybersecurity data limited to “defensive” content, filtering out exploit development, reverse engineering and penetration testing material, according to Reuters. Tencent’s launch announcement promised weights and system cards “in the coming days” with “a focus on cybersecurity and biological safety.” It published no before-and-after measurement.

Alibaba (Qwen4-Max, November 16). Alibaba combined data filtering with a release decision. Its launch material says it kept training data for defensive tasks such as vulnerability verification and remediation and removed data meant to enhance attack capability. It also shipped the flagship without weights, which Bloomberg highlights as the more consequential step: nobody can download and modify the full model. Smaller Qwen4 models are promised as open source without a date. Our Qwen4-Max coverage has the details.

Bloomberg also notes a fourth approach from Z.AI, which ran security reviews and alerted organizations to vulnerabilities before making GLM-5.3 open-weight in August, rather than stripping capability.

The regulatory signal behind it

The Ministry of Commerce and the Ministry of Science and Technology published a draft revision of China’s catalog of banned and restricted export technologies in mid-October. Per Bloomberg, it proposes an export ban on AI tools capable of attacking critical national infrastructure, government networks or military targets, and says large language models deemed “world-leading” could face restrictions on overseas deployment.

Secondary coverage of the draft, including Z Media’s summary of reporting by The Information and The Decoder, describes 23 AI areas spanning large models, agents, compute clusters and data processing. Three provisions are relevant to open-weight users:

  • Autonomous tools that write or alter malware to hit infrastructure, military or government targets would be banned from export outright.
  • Core model technology such as weights and training methods would be controlled once a system crosses thresholds for size, training compute or global ranking.
  • Publishing restricted technology as open source would qualify for automatic licensing once paperwork is filed, technology that is already public is exempt, and access through APIs, apps, websites or SaaS stays outside the controls.

So the draft, as described, does not prohibit open-weight releases of large models. It creates a registration step and a legal category. That reading is consistent with what the labs are doing: Moonshot and Tencent still release weights, but with the capability most likely to fall into the banned category reduced first. The draft is not final, and the thresholds may change.

Bloomberg adds a second signal: in early November an official at the National Development and Reform Commission said China is accelerating new safety standards and ethical guidelines for AI developers and intends to embed security measures during development.

Why open weights are the pressure point

Hosted models can be monitored, rate-limited and cut off. Open weights cannot: an attacker runs them locally, removes refusals and fine-tunes on whatever was filtered out. Bloomberg cites CrowdStrike research from October finding that hackers used DeepSeek’s model in attacks on organizations in 14 countries, and recalls that Moonshot’s Kimi K3 broke out of a test environment at the UK’s AI Security Institute. We covered one such case last week, where leaked transcripts showed a jailbroken DeepSeek agent attacking a European telecom for two weeks.

“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,” Tencent chief AI scientist Yao Shunyu said at a Hong Kong event this month, per Bloomberg.

US labs have handled the same problem with access control rather than data removal: Anthropic keeps its most cyber-capable Mythos models with vetted organizations, and OpenAI restricts GPT-6 Cyber to approved security customers. That option only exists for closed models. A lab committed to open weights has to change the model itself, which is why training-data filtering has become the Chinese default and why Alibaba’s decision to withhold its flagship stands out.

There is also a commercial reading. Charles Li, chief analyst at TeamT5, told Bloomberg that Beijing “increasingly views these capabilities as strategic assets that warrant tighter export controls” as Chinese frontier models close the gap with US ones. In that framing, the filtering is as much about keeping a capability at home as about global safety.

How much capability is actually removed

Less than the framing suggests, on the one number available. Moonshot’s shipping model still solves 81% of its own hard exploit benchmark. The filtering cost 9.5 points, and the result remains far above the previous generation. A filtered model that is strong at general software engineering retains most of what offensive work requires: reading unfamiliar code, reasoning about memory and state, and iterating against a target.

Xiaojian Li, founder of AI safety firm Fangcun AI, told Bloomberg that stripping offensive material from training is “a good balance between cost and effectiveness, but it’s not perfect,” and that “there is no such method now that can completely eliminate models’ risk.”

Three gaps limit what outsiders can conclude:

  • No independent measurement. Alibaba cites an internal expert review. Tencent cites internal testing. Only Moonshot gave a figure, on its own benchmark. None of the three has published a third-party offensive evaluation of the shipped model against an unfiltered checkpoint.
  • Fine-tuning reversibility. Data filtering removes examples, not the underlying coding ability. How quickly a modest fine-tune on public exploit write-ups restores the lost points on open weights is the question that decides whether the approach works, and no lab has reported testing it.
  • Defensive cost. Exploit development, reverse engineering and penetration testing are also what defenders do. Security teams that use open models for red-team automation or vulnerability triage should expect weaker results from these releases and test before upgrading.

What to watch

  • The final export catalog. Whether the open-source licensing route survives, and where the size and ranking thresholds land, determines whether flagship weights keep shipping.
  • Hy4’s system card. Tencent promised one with the weights. A measured before-and-after, like Moonshot’s, would make its claim checkable.
  • The smaller Qwen4 models. If they arrive as open weights with the same filtering, Alibaba’s flagship-closed, small-open split becomes the template.
  • DeepSeek. It is the lab named in the attack reporting and is absent from Bloomberg’s list of firms that have cut capability. Its next release will show whether the pattern is industry-wide.
Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Agents

How to Secure AI Agents With Google ADK

Learn how to secure your autonomous workflows and prevent unauthorized actions using Google ADK's hardware-backed tool binding and execution logs.

Ai Engineering

Microsoft's Ignite model drop: MAI Cyber 2 Flash, Code 2, Realtime 1 and Decision 1.1 aim at price, not the frontier

At Ignite on November 17, Microsoft AI announced seven first-party models, from a 1M-context security model priced at $0.50 per million output tokens to coding models in GitHub Copilot and an OpenAI-compatible speech-to-speech model at $0.01 a minute. Here is what each does, where it runs and what the benchmark fine print says.

Ai Engineering

Hugging Face hands LLM routing to OpenRouter: failover by default, custom provider keys gone

On November 16, Hugging Face replaced the custom router behind Inference Providers with OpenRouter's for all language models. Requests now fail over across providers inside one call, pricing is unchanged, and bring-your-own provider keys stop working for chat models.

Ai Engineering

Qwen4-Max: Alibaba's 2.4T closed model cuts prices and claims a 62% cheaper task, but trails US rivals on independent tests

Alibaba released Qwen4-Max on November 16 with a new Gated DeltaNet plus sparse-attention architecture, 71B active parameters, 1M context and $1.50/$4.50 pricing. Its own Work360 benchmark shows a 62% lower cost per task; Artificial Analysis puts it at 43, behind Kimi K3.1 Pro.

Ai Engineering

US closes the cloud loophole: new BIS rule licenses AI compute for China-linked customers and five diversion hubs

BIS's November 16 interim final rule requires a license for advanced AI chips going to any data center that will give China- or Macau-headquartered firms more than 12 million aggregate TPP, adds license requirements for Brazil, Indonesia, Malaysia, Singapore and Thailand, and formally buries the AI Diffusion Rule.