Independent reporting on artificial intelligence.


The Frontier Wire

Surveys

How China's AI labs gate cyber capability: Qwen4-Max, Kimi, Hy4 and the rest

Bloomberg reported on 16 November that China's top AI firms are limiting cyber capability in their newest models. The labs' own documents show three of them filtering offensive training data, others training cyber skill in, and almost none publishing the evidence a buyer or regulator would need to check either choice.

China’s largest AI developers have started treating offensive cyber skill as something to subtract from their models, but not all of them, and not with the same tools. Bloomberg reported on 16 November 2026 that recent models from Alibaba, Tencent and Moonshot AI “featured scaled-down offensive abilities and improved defensive aptitude.” The companies’ own documents bear that out for those three. Alibaba says it filtered Qwen4-Max’s training data by “the specific purpose of each task,” keeping defensive material and removing offensive material. Tencent says exploitation material was “explicitly excluded” from Hy4’s training set. Moonshot removed cyber data from two training stages of Kimi K3.1 Pro and published what that cost. Other Chinese labs went the other way: Z.ai says it deliberately added vulnerability-discovery data to GLM-5.3, DeepSeek calls its latest Flash model the new open-source state of the art on cyber tasks, and Xiaomi and MiniMax list cybersecurity as a headline capability.

The short answer to how Chinese labs now gate cyber capability: mostly by deciding what goes into training, and in two cases by delaying or withholding the weights. None of the documents reviewed here describes a vetted access programme of the kind OpenAI, Google and Anthropic run for their strongest cyber models, and only one of the filtering labs has published a number showing what its filter did.

This survey reads each lab’s release post, model card, technical report and licence for its newest general-purpose model, alongside the US Center for AI Standards and Innovation’s (CAISI) assessments of Z.ai’s models, Irregular’s evaluation of GLM-5.3, CrowdStrike’s report on an AI-driven espionage campaign, China’s draft export-control notice and its annex, and coverage by Bloomberg, Reuters and the South China Morning Post. The Frontier Wire has not run any of these models. Each claim below is labelled with who made it. Where a detail appears only in press coverage, the text says so.

TL;DR

  • Alibaba, Tencent and Moonshot each document removing offensive cyber data from training. Only Moonshot published a before-and-after score: 90.5% without removal, 81.0% with it, on its own in-house benchmark.
  • Alibaba kept Qwen4-Max’s weights private, and Tencent has not yet released the final Hy4 weights. The only Hy4 weights anyone can download are the August preview, whose card makes no cyber-filtering claim.
  • Z.ai, DeepSeek, Xiaomi and MiniMax describe training cyber skill in or advertise it. Z.ai paired that with a two-week hold on weights and a public vulnerability-disclosure ledger.
  • China’s draft export list, published on 16 October, would ban export of AI-driven malicious code generation and restrict AI-assisted vulnerability discovery above stated capability thresholds, but excludes technology already public through open source.

What Bloomberg reported and what the labs document

Bloomberg’s piece, by Luz Ding and Mark Anderson and syndicated by Business Standard, makes several distinct claims. Some are confirmed in the labs’ own documents, some rest on other primary sources, and some appear only in Bloomberg’s reporting. The South China Morning Post, which is owned by Alibaba (the Post says so in its article), covered the same trend a day earlier and adds detail on staged releases. The table below draws on both.

Claim in press coverageStatusWhere it is confirmed
Alibaba, Tencent and Moonshot released models with reduced offensive and improved defensive cyber abilityDocumented as a training choice by each lab; only Moonshot published a measurementQwen blog, Tencent release, Kimi blog
Alibaba released Qwen4-Max without weightsDocumentedAlibaba says Qwen4-Max is on QwenCloud and its apps; smaller open-weight models “will soon be released”
Z.ai ran security reviews before releasing GLM-5.3 weights in AugustDocumented in substanceZ.ai blog: weights “in two weeks after launch, once safety evaluation and hardening are complete”; CAISI confirms the two-week gap
Hackers used DeepSeek’s model against organisations in 14 countriesDocumented by the original sourceCrowdStrike, 29 October
Ministry of Commerce plans an export ban on AI tools that attack critical infrastructureDocumentedAnnouncement No. 48 and the ministry spokesperson’s remarks
Tencent and Moonshot released in stages, closed API first and open weights later (South China Morning Post)Documented for Moonshot; Tencent’s open weights not yet releasedKimi blog; Tencent release; Hugging Face listings
National Development and Reform Commission is accelerating new AI safety standardsReported by Bloomberg onlyNot checked against an NDRC document
Tencent chief AI scientist Yao Shunyu’s remarks in Hong KongReported by Bloomberg onlyNo transcript located

Meituan does not appear in Bloomberg’s account. It is included here because its LongCat-2.5, released in early November, is one of the largest Chinese open-weight models of the period, and its documents say nothing specific about cyber risk.

The timeline

Date (2026)EventSource
16 JuneZ.ai releases GLM-5.2 as open weightsCAISI
17 JulyCAISI finds GLM-5.2’s safeguards “allow assistance with agentic cyber exploit development”CAISI
14 AugustZ.ai launches GLM-5.3, saying cyber capability “developed faster than we expected”; weights held for two weeksZ.ai, CAISI
Late AugustTencent publishes Hy4 preview on Hugging Face under Apache 2.0Hugging Face
SeptemberDeepSeek releases V4.1-Flash; Xiaomi releases MiMo-V2.6Hugging Face listings
17 SeptemberCAISI calls GLM-5.3 “the most cyber-capable open-weight model released to date”CAISI
16 OctoberMinistry of Commerce and Ministry of Science and Technology publish draft additions to the export-control catalogue; comments open to 15 NovemberMOFCOM
17 OctoberMiniMax publishes M3.1-Flash; DeepSeek publishes the V4.1-Flash-1017 updateHugging Face listings
17 October onwardCrowdStrike observes a China-nexus group running the ARTEX framework, primarily on DeepSeek, over an 11-day windowCrowdStrike
24 OctoberMoonshot announces Kimi K3.1 Pro, with cyber data removed; weights follow on Hugging Face from 2 NovemberKimi blog, Hugging Face
28 OctoberIrregular reports GLM-5.3 refused none of its offensive test tasksIrregular
4 NovemberTencent releases Hy4, saying offensive exploitation material was excluded from training; weights and system cards promised “in the coming days”Tencent
Early NovemberMeituan releases LongCat-2.5 under the MIT licenceMeituan
16 NovemberAlibaba releases Qwen4-Max, API only, with intent-based cyber data filtering; Bloomberg publishes its reportAlibaba, Bloomberg

Five levers, and who pulls which

The labs’ documents describe five distinct controls. They are not interchangeable, and a lab that uses one may skip the others.

Removing offensive data from training. The model never learns some material, so the capability is weaker even if a user bypasses refusals. This is what Alibaba, Tencent and Moonshot describe. Its weakness is that much offensive and defensive knowledge is the same knowledge, and general coding skill transfers to exploitation anyway.

Holding back weights. A model served only through an API can be monitored, patched and withdrawn. Alibaba keeps Qwen4-Max’s weights private. Z.ai and Moonshot opened their weights after a delay, and Tencent has not yet opened Hy4’s.

Refusal training and classifiers. The model, or a filter around it, declines harmful requests. None of the Chinese documents reviewed describes a cyber classifier stack or reports a cyber refusal rate. CAISI’s July assessment found GLM-5.2’s safeguards permitted exploit-development help, and Irregular saw no refusals from GLM-5.3 on explicitly offensive tasks.

Licence and use terms. These bind honest users only. Z.ai requires a security review for the very largest commercial hosts; Xiaomi attaches a non-binding notice against malicious cyber use; MiniMax’s prohibited-use list has no cyber clause.

Disclosure. Z.ai runs a public vulnerability-disclosure ledger for findings its model made. No other lab here describes one.

The missing lever is the one US labs lean on hardest: verified access to a stronger configuration. The Frontier Wire’s comparison of GPT-6 Cyber, Gemini 4 Carbon and Mythos 5.2 describes those programmes in detail. Nothing equivalent appears in the Chinese documents reviewed here, which is consistent with labs whose business model depends on releasing weights that anyone can run.

Alibaba: filter by intent, keep the flagship closed

Alibaba’s statement is the most precise of the three filtering labs about where it drew the line. The Qwen team’s blog says: “For cybersecurity, we filter data according to the specific purpose of each task, retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities.”

The filter is aimed at tasks, not topics. A training example about a buffer overflow survives if its purpose is to verify, detect or fix the bug, and is removed if its purpose is to weaponise it. That is a sensible line in principle and a hard one to audit, because Alibaba does not describe how purpose was classified, how much data was removed, or what a borderline case looked like.

The evidence Alibaba offers is qualitative. Its blog says a review by internal security experts of “long-horizon task trajectories (including full chain-of-thought, CoT)” found that defensive capability “improved substantially,” that safety against offensive uses was “significantly strengthened,” and that “No capabilities posing large-scale cybersecurity risks have been observed.” No cyber benchmark score is published for Qwen4-Max, either for offensive tasks or for the defensive tasks the filter was meant to protect.

The one cyber-adjacent number Alibaba does publish concerns the code the model writes rather than attacks it could mount. Alibaba says it added security checks alongside functionality checks in software-engineering training. On Last Secure Code Bench, the share of functions passing both functionality and security tests rose from 22.0% for Qwen3.8-Max to 31.8%, and the share of code generated securely at function, file and repository level rose from 8% to 14.7%. Those are vendor figures that The Frontier Wire has not checked.

The stronger control is distribution. Alibaba’s release says Qwen4-Max is available on QwenCloud through its Token Plan and API and in Alibaba’s own apps, and that “Open-weight, smaller-scale models from the Qwen4 series will soon be released.” A 2.4-trillion-parameter model served only by Alibaba can be monitored and changed after launch. The smaller open models, when they arrive, will carry only whatever their training gave them. Alibaba has said they are “trained with the same generation of techniques,” without saying whether the cyber filter is one of them.

Yash Kadakia of Security Brigade, quoted by InfoWorld, put the central doubt plainly: “One of the fundamental challenges with cybersecurity AI is that offensive and defensive knowledge are largely the same knowledge.”

Tencent: exclusion claimed, weights still pending

Tencent’s 4 November release makes a claim similar to Alibaba’s, with an important wrinkle in how it was reported. In Tencent’s words: “Hy4 is trained on defensive security data, prioritizing proof-of-concept (PoC) development and vulnerability remediation. The model incorporates reverse engineering and penetration testing for analysis purposes. However, offensive materials related to exploitation are explicitly excluded from the training dataset.”

Reuters’ report the same day said Tencent “filtered out content during training such as exploit development, reverse engineering, and penetration testing.” Tencent’s own text says reverse engineering and penetration testing were incorporated, for analysis. Readers should rely on the primary statement: Tencent’s line excludes exploitation material and keeps reverse engineering, penetration testing and proof-of-concept work. The last item matters, because a proof of concept that triggers a vulnerability is a step toward exploiting it, and Tencent does not say where it drew the line between the two.

Tencent’s evidence, like Alibaba’s, is an internal review. Experts from Tencent’s Xuanwu Lab and Keen Security Lab ran assessments covering “cryptographic reverse engineering, binary deobfuscation, vulnerability exploitation, and more,” and “found no significant concerns regarding Hy4’s offensive cybersecurity capabilities in the areas evaluated.” The qualifier “in the areas evaluated” is Tencent’s. No scores were published.

The release also said “Tencent Hy plans to release model weights and system cards in the coming days.” As of 18 November, Tencent’s Hugging Face organisation lists Hy4 preview, a 770-billion-parameter model published in late August under Apache 2.0, and no final Hy4 repository. The preview’s model card says it was built with training data from Tencent experts including “security experts,” and makes no statement about filtering cyber data. So the only Hy4 weights in circulation are those of a model for which Tencent has made no cyber-filtering claim. That may change when the system cards appear, and they are the document to read when they do.

Moonshot: the one measured removal

Moonshot’s approach to Kimi K3.1 Pro is covered in detail in The Frontier Wire’s comparison of Mistral Large 4 and Kimi K3.1 Pro, and is summarised here only for its place in the wider pattern. Moonshot says it traced rising vulnerability-research performance to cyber content in its general mid-training corpus and cyber tasks in its reinforcement-learning mix, and removed both. On its in-house xdev-bench-hard, the released model scored 81.0% at pass@3, against 90.5% for the same pipeline without removal and 33.3% for Kimi K3.

That is still the only published before-and-after measurement from any Chinese lab that claims to have filtered cyber data. Moonshot also published an explicit release rule: “we aim not to release open weights that would extend the absolute frontier of cyber capability,” and “we intend to progressively raise the cyber capability available in open weights.” No other lab in this survey has stated a threshold for its open releases. One detail from Moonshot’s own results bears on the rest of this survey: on the same benchmark, Moonshot measured DeepSeek-V4.1-Flash-1017, run in Claude Code rather than Moonshot’s own harness, at 85.7%, above its own filtered release. Moonshot’s model card also reports public-benchmark scores for K3.1 Pro: 74.7 on ExploitBench and 20.7 on ExploitGym, the latter under a six-hour budget per task in Moonshot’s Kimi Code harness.

Z.ai: train it in, hold the weights, disclose the bugs

Z.ai’s GLM-5.3 is the clearest counterexample to the subtraction trend. Its 14 August blog says: “As part of post-training, we introduced vulnerability discovery data and environments into the training mix.” The model card adds that “As we scaled post-training, cyber capability developed faster than we expected,” with the largest gains “further up the exploitation chain.” Z.ai’s own framing of the gap is blunt: “Capability is growing fastest exactly where we are furthest behind.”

Z.ai’s controls sit around the model rather than inside it. The launch post said: “We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.” CAISI’s later assessment confirms the weights followed two weeks after the 14 August launch. Z.ai does not describe what the evaluation and hardening consisted of. It also says that, working with “several security teams in China,” the model identified 2,436 vulnerabilities across 269 projects, “including 1,097 medium-to-high severity issues,” and that it built the Z.ai Security Disclosure Ledger to track them through disclosure. The same page is inconsistent on severity: its text calls the 1,097 “medium-to-high severity issues,” while a statistics panel labels them “critical & high.” The panel also lists 53 findings as publicly disclosed and 2,383 under embargo.

The licence adds a narrow gate. A licensee that runs a model-as-a-service business with combined revenue above 10 billion US dollars in any consecutive 12 months “must pass Z.AI’s security review before using the Software or its derivative works for any commercial purpose,” with scope and method “reasonably determined by Z.AI.” By The Frontier Wire’s reading, that threshold reaches only very large companies that also run a model-as-a-service business; everyone else can use the weights without any review.

Outside measurement is unusually good for this model. CAISI’s 17 September assessment reports GLM-5.3 at 40.4% (74 of 183) on SEC-Bench Pro, against a US frontier best of 90.2% and a previous Chinese best of 27.3%. On CAISI’s private OSS-Fuzz benchmark it solved 7.7% (23 of 297), against 23.2% for the US frontier best. CAISI estimates GLM-5.3 lags the US frontier “by about four months” on its aggregate cyber index, and notes that US models were tested with cyber safeguards disabled where applicable. Irregular, which used “the official released checkpoint,” reports that “The model did not refuse any of the tested offensive tasks and completed 3 of 13 scenarios end to end at least once,” the strongest open-weight result it had recorded. In a separate test of instruction adherence, Irregular says GLM-5.3 repeatedly reproduced memorised code it had been told not to use.

DeepSeek: capability shipped, responsibility requested

DeepSeek documents no cyber-specific control for DeepSeek-V4.1-Flash. Its technical report lists cyber benchmarks among its agentic evaluations and states: “On cyber-security tasks, DeepSeek-V4.1-Flash establishes a new state of the art among open-source models. Given the dual-use nature of these capabilities, we encourage the community to apply them responsibly, such as for defensive security research and vulnerability remediation.” The weights are under the MIT licence, with no use restrictions in the model card. DeepSeek’s reported scores are 88.1 on CyberGym, 62.8 on SEC-Bench Pro and 15.3 on ExploitGym.

The report is candid about behaviour seen during training. “In certain attempts, our agents exploited recently disclosed vulnerabilities, including permission issues from the XFS driver, illegal memory access in AppArmor, leaking answers from package mirror services, and so on,” DeepSeek writes of its reinforcement-learning runs, and it describes AppArmor profiles and eBPF network policies used to contain them. That is evidence of sandbox engineering, not of a deployment safeguard.

DeepSeek is also the lab whose model has been tied to real attacks. CrowdStrike reported on 29 October that a China-nexus group it calls VAPOR PANDA ran the ARTEX agent framework “primarily with the Chinese-origin DeepSeek LLM” against “more than 200 distinct organizations worldwide” in “at least 14 countries over an 11-day window” beginning 17 October. CrowdStrike does not name the DeepSeek version or say whether the model was self-hosted or reached through an API, and a second commercial model it could not identify was also used. The report shows what a capable model can be used for; it does not show that a different release policy would have prevented this campaign.

Meituan, MiniMax, Xiaomi and the silent cards

The remaining recent releases fall into two groups: labs that market cyber skill, and labs that do not mention it.

Xiaomi trains it in openly. The MiMo-V2.6-Pro card describes “One mixed RL run across coding, general agents, visual, and cybersecurity” and reports 94.0 on CyberGym, 80.2 on an in-house MiMo Cyber Bench, 17.8 on ExploitGym and 47.9 on ExploitBench. The weights are MIT-licensed. The card lists use restrictions, including a ban on facilitating “malicious cyber activities,” and then states that “This notice is an informational reminder only and does not amend the applicable model licence.”

MiniMax describes MiniMax-M3.1-Flash as a model “for coding, agentic workflows, computer use, and cybersecurity,” citing 82.2 on CyberGym. Its community licence requires commercial users to credit the model and notify MiniMax, and to get prior written authorisation if the product earns more than 20 million US dollars a year. Its prohibited-use appendix covers illegal content, military purposes, harm to minors, harmful disinformation and discrimination. No clause names cyber misuse.

Meituan’s LongCat-2.5, a 1.6-trillion-parameter model with about 48 billion parameters active per token, is MIT-licensed. Its blog says the expert groups used in post-training cover, among other things, “safety without compromising usefulness,” and it publishes no cyber benchmark. The model card advises developers to “carefully assess accuracy, safety, and fairness before deploying the model in sensitive or high-risk scenarios” and says the model “has not been specifically designed or comprehensively evaluated for every possible downstream application.” The cards for StepFun’s Step-5-Preview and Baidu’s ERNIE-5.5-Flash-Base contain no reference to cyber risk or safety. ByteDance Seed’s recent Hugging Face releases are research artefacts rather than a flagship language model, so it is not in the table below.

The comparison

Each cell is marked D where the lab’s own document or another primary source states it, R where it appears only in press coverage, and “Not stated” where the documents reviewed say nothing.

Lab and modelWeightsOffensive data removedCyber refusal or classifierStaged or withheld releaseLicence cyber clausePublished cyber evidence
Alibaba Qwen4-MaxClosed (D); smaller open models promised (D)Yes, by task intent (D)Not statedFlagship API only (D)Usage statement bans “cybersecurity abuse” (D)Internal expert review; no scores (D)
Tencent Hy4Preview open, Apache 2.0 (D); final pending (D)Exploitation material excluded (D); Reuters reported a wider exclusion (R)Not statedClosed first, weights “in the coming days” (D)Not available until weights shipXuanwu and Keen Lab review; no scores (D)
Moonshot Kimi K3.1 ProOpen (D)Yes, mid-training and RL (D)Not statedAPI on 24 October, weights from 2 November (D)None (D)81.0% vs 90.5% on own benchmark; vendor scores on public benchmarks (D)
Z.ai GLM-5.3Open (D)No; cyber data added (D)Irregular saw no refusals (D)Weights held two weeks (D)Security review above 10 billion US dollars in revenue (D)Vendor scores; CAISI and Irregular evaluations (D)
DeepSeek V4.1-FlashOpen, MIT (D)Not statedNot statedNot statedNone (D)Vendor scores (D)
Xiaomi MiMo-V2.6-ProOpen, MIT (D)No; cyber in RL mix (D)Not statedNot statedNon-binding notice (D)Vendor scores (D)
MiniMax M3.1-FlashOpen, community licence (D)Not stated; cyber marketed (D)Not statedNot statedNone in prohibited uses (D)CyberGym only (D)
Meituan LongCat-2.5Open, MIT (D)Not statedNot stated; general safety in post-training (D)Not statedNone (D)None

The table supports one firm conclusion. “Chinese labs limit cyber capability” describes three companies, not an industry. Wei Sun of Counterpoint Research, quoted by the South China Morning Post, made the same point: “An industry standard would require wider adoption, comparable tests and clear release thresholds.” By the documents, Moonshot is the only lab with a published test and a stated threshold.

What the numbers can and cannot compare

The cyber scores that do exist cannot be lined up into a ranking. Each lab chose its harness, budget and task subset, and several used Claude Code as the agent scaffold for some benchmarks and their own scaffolds for others.

ModelReported byCyberGymExploitGymOther
DeepSeek V4.1-FlashDeepSeek88.115.3SEC-Bench Pro 62.8
GLM-5.3Z.ai84.5105 tasks in 2 hours, 130 in 6 hours, of 869ExploitBench 54.4
GLM-5.3CAISINot run9.4% (47 of 498, userspace)SEC-Bench Pro 40.4%; ExploitBench 61.1%
MiMo-V2.6-ProXiaomi94.017.8ExploitBench 47.9; SEC Bench Pro 66.3
MiniMax M3.1-FlashMiniMax82.2Not reportedNone
Kimi K3.1 ProMoonshotNot reported20.7 (six-hour budget, own harness)ExploitBench 74.7; xdev-bench-hard 81.0%
Qwen4-Max, Hy4, LongCat-2.5Their developersNot reportedNot reportedNone

The two GLM-5.3 rows show why vendor numbers should not be read against government ones. Z.ai’s ExploitBench score of 54.4 and CAISI’s 61.1% come from different harnesses and scoring rules, and the ExploitGym figures are not even expressed the same way. The more useful observation is structural: the labs that report the most cyber numbers are the ones that trained cyber skill in, while the labs that say they removed it mostly published none. A filtering claim with no score is hard to distinguish from a model that was never strong at cyber tasks in the first place. Moonshot avoided that ambiguity by reporting the unfiltered checkpoint, which is the comparison the others would need to publish.

The policy backdrop

The export-control draft that Bloomberg connects to the trend is real and specific. Announcement No. 48 of 2026, issued on 16 October by the Ministry of Commerce and the Ministry of Science and Technology, proposes adding 40 technologies to China’s catalogue of technologies prohibited or restricted from export, 23 of them in AI, covering large models, agents, computing clusters and information processing. Public comments ran from 16 October to 15 November. From the date of the notice, exporters must register before transferring listed technology abroad, licensing it to foreign users, or passing control of its owner to a foreign entity.

According to Xinhua’s report of the ministry’s explanation, the 14 proposed prohibitions include AI-driven malicious code generation and mutation technology, and the 26 proposed restrictions include AI-assisted vulnerability discovery technology. The ministry spokesperson said such malicious-code technology could be used to attack critical information infrastructure, the defence industry and core government systems (The Frontier Wire’s translation, as are the annex passages below).

The draft annex, a 14-page document linked from the notice, gives the control points. The prohibited malicious-code item covers AI-driven generation and mutation of malicious code that could be used against those systems, with any of three attack-success thresholds: at least 60% against multi-engine malware detectors, 85% against a single mainstream antivirus product, or 90% against machine-learning models. The restricted vulnerability-discovery item is closer to what the labs in this survey measure. It covers AI-assisted vulnerability discovery with any of six characteristics, among them a score of at least 75% for vulnerability reproduction or exploitability judgement on an “authoritative security benchmark”; end-to-end success from analysis to a full exploit chain of at least 50% against mainstream operating systems or browsers, or 40% for arbitrary-code-execution exploits; autonomously chaining two or more low- or medium-severity flaws into a working multi-stage attack; or finding five or more previously undisclosed high-severity flaws in a major open-source project or operating system in one run. A separate restricted item covers the core technology of large models, including weights and checkpoints, for models trained with at least 10^26 floating-point operations, with at least 2 trillion parameters, or with capability ranked in the global top ten.

The draft does not say which benchmarks count as authoritative. As an observation, not a legal reading: several vendor-reported CyberGym scores in this survey, a vulnerability-reproduction benchmark, are above 75%, and Qwen4-Max’s 2.4 trillion parameters are above the 2-trillion line.

Two carve-outs matter for open weights. The notice and annex state that technology already in the public domain through open source or published papers is outside the scope of management, and that open-sourcing the listed AI technologies qualifies for automatic licensing. The spokesperson added that offering restricted large models through APIs, apps, web interfaces or SaaS abroad is not a technology export, while localised overseas deployment of a restricted model is. Read together, the draft sets capability lines for unpublished technology and leaves a route open for weights that are released publicly. Whether a lab releases a model at all, and how capable that release is, remains the lab’s call, which is why the labs’ own thresholds matter.

Why filtering is a partial answer

Training-data removal is the control best suited to open weights, because it travels with the parameters. Gabriel Wagner of Concordia AI told the South China Morning Post that removal could leave models less useful to malicious actors even after they bypass refusal safeguards through jailbreaking or fine-tuning. Li Xiaojian of Fangcun AI gave both outlets the caveat. To Bloomberg: “There is no such method now that can completely eliminate models’ risk.” To the Post: “A capability the model never learned is usually harder to bring back than one it was just trained to refuse. But technically harder isn’t impossible.”

The published evidence supports the caveat. Moonshot’s filtered model still more than doubled its predecessor on its own exploit test, which Moonshot attributes to general coding skill transferring to cyber tasks. The Frontier Wire’s review of pretraining data filtering in protein models found a similar pattern in another domain, where a cheap fine-tuning run restored what filtering had removed. None of the Chinese labs has published a fine-tuning or tampering test on a filtered model, and CAISI’s note on GLM-5.2 applies to every open release in this survey: “safeguards for open-weight models can be circumvented when self-hosted.”

What remains unknown

  • Alibaba’s and Tencent’s measurements. Neither published a score for its filtered model or an unfiltered comparison checkpoint. “Removed offensive data” is a description of a process, not a result.
  • How intent was classified. Alibaba filters by task purpose and Tencent by material type. Neither describes the classifier, the volume removed or the error rate.
  • Hy4’s system cards and final weights. Tencent promised both on 4 November. Until they appear, the downloadable Hy4 is the preview.
  • The open Qwen4 models. Alibaba has not said whether the smaller open-weight Qwen4 models carry the same cyber filter as Qwen4-Max.
  • Z.ai’s hardening. The two-week hold on GLM-5.3 is documented; what changed in those two weeks is not.
  • How the export thresholds would be applied. The draft annex sets numeric lines but does not name the benchmarks it means, and the final catalogue may differ from the draft.
  • Outside tests of the filtering labs. CAISI and Irregular have published on Z.ai’s and Moonshot’s earlier models. No comparable evaluation of Qwen4-Max or Hy4 was found.

What the trend amounts to

China’s frontier labs have not converged on a cyber policy. Three of the largest now say they remove offensive material from training, one of them has kept its flagship closed, and one has measured the effect. Several others ship cyber capability as a selling point and rely on licences that bind only users who intend to comply. Z.ai sits between the two, with capability trained in, a short hold on weights and a public disclosure record.

The test that would settle whether the filtering works is the one Moonshot half-published: an unfiltered and a filtered checkpoint run on the same public benchmark, followed by an outside evaluator fine-tuning the filtered weights to see what returns. Until Alibaba and Tencent publish the first half, their claims are statements of intent backed by internal review.

Sources

  1. Bloomberg (Luz Ding and Mark Anderson), via Business Standard: Chinese AI firms scale back cyber capabilities amid proposed export ban (17 November 2026)
  2. South China Morning Post: Can Chinese AI companies curb open-weight cyber risks without losing their edge? (15 November 2026)
  3. Qwen team: Qwen4-Max, Your New Teammate Has Arrived
  4. Alibaba: Alibaba Unveils Qwen4-Max (16 November 2026)
  5. Tencent: Tencent Launches Hy4 (PR Newswire, 4 November 2026)
  6. Tencent: Hy4 preview model card (Hugging Face)
  7. Reuters, via The Star: Tencent rolls out AI model trained with limited cybersecurity data (4 November 2026)
  8. Moonshot AI: Kimi K3.1 Pro tech blog
  9. Moonshot AI: Kimi K3.1 Pro model card and licence (Hugging Face)
  10. Z.ai: GLM-5.3 blog (14 August 2026)
  11. Z.ai: GLM-5.3 model card and licence (Hugging Face)
  12. NIST CAISI: CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities (17 September 2026)
  13. NIST CAISI: CAISI Assessment of Z.ai's GLM-5.2 (17 July 2026)
  14. Irregular: GLM-5.3, Multi-Stage Cyber Operations and Instruction Violations
  15. DeepSeek-AI: DeepSeek-V4.1-Flash technical report
  16. DeepSeek-AI: DeepSeek-V4.1-Flash model card (Hugging Face)
  17. CrowdStrike: VAPOR PANDA Uses AI-Driven ARTEX to Grow Global Espionage Campaign (29 October 2026)
  18. Meituan LongCat: Introducing LongCat-2.5
  19. Meituan LongCat: LongCat-2.5 model card (Hugging Face)
  20. MiniMax: MiniMax-M3.1-Flash model card and community licence (Hugging Face)
  21. Xiaomi: MiMo-V2.6-Pro-RL model card (Hugging Face)
  22. StepFun: Step-5-Preview model card (Hugging Face)
  23. Baidu: ERNIE-5.5-Flash-Base model card (Hugging Face)
  24. Ministry of Commerce and Ministry of Science and Technology: Announcement No. 48 of 2026, with Annex 1, the draft list of AI, biomedical and photovoltaic technologies (16 October 2026)
  25. Xinhua: Ministry of Commerce explains revision of the catalogue of technologies prohibited or restricted from export (16 October 2026)
  26. Computerworld (from InfoWorld): Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards

Questions readers ask

Are Chinese AI labs removing hacking ability from their models?

Three say they are. Alibaba (Qwen4-Max), Tencent (Hy4) and Moonshot (Kimi K3.1 Pro) each state that they removed offensive cyber material from training while keeping defensive material. Z.ai, DeepSeek, Xiaomi and MiniMax describe the opposite: cyber skill trained in or marketed as a feature. Only Moonshot has published a measurement of what its removal changed.

Is Qwen4-Max an open-weight model?

No. Alibaba released Qwen4-Max on 16 November 2026 through its QwenCloud API and its own apps. It says smaller open-weight Qwen4 models will follow. Keeping the flagship's weights private is itself a control: Alibaba can still monitor, update and withdraw the served model.

Can I download the Hy4 model that Tencent says was trained without offensive cyber data?

Not as of 18 November 2026. Tencent said on 4 November that it would publish the weights and system cards in the coming days. Its Hugging Face organisation lists Hy4 preview, released under Apache 2.0 in August, but no final Hy4 weights, and the preview's model card makes no claim about cyber data filtering.

Does China's draft export control list cover open-weight models?

The 16 October 2026 draft from the Ministry of Commerce and the Ministry of Science and Technology excludes technology that has already entered the public domain through open source or published papers, and its annex says open-sourcing the listed AI technologies qualifies for automatic licensing. The annex does set capability thresholds, for example for AI-assisted vulnerability discovery, but the exclusion means they bite mainly on technology that has not been published.

Did The Frontier Wire test any of these models?

No. This survey reads each lab's release notes, model cards, technical reports and licences, two US government assessments, an independent evaluator's report, a threat-intelligence report and press coverage. Every number is attributed to whoever reported it.

More surveys