Part of the AI Regulation News hub.
China's market regulator published a trade-secret enforcement case finding that a vertical-domain AI model's prompt templates, review rules and annotation specifications can independently constitute trade secrets, apart from source code
The enforcement is ordinary: an ex-employee took materials to a company he controlled, and was fined RMB 350,000 under the Anti-Unfair Competition Law. What is not ordinary is what the regulator counted as the secret. SAMR says past AI intellectual-property work was mostly confined to establishing rights in source code, and that this case breaks the industry's code complex.
Bottom line: An administrative trade-secret penalty, not a court judgment and not a new AI rule. The Hangzhou Municipal Market Regulation Bureau ordered Sun to stop the conduct and fined him RMB 350,000 on 28 May 2026 under the Anti-Unfair Competition Law. SAMR published it nationally on 20 August 2026 as one of six typical cases.
Who this affects: AI developers whose competitive edge sits in prompt design, review logic and annotation practice rather than in model weights or code; employers of algorithm staff in China; and anyone relying on confidentiality agreements to protect non-code AI know-how.
What to do: Identify which of your AI assets are natural-language or operational rather than code, and check whether your confidentiality controls actually reach them. SAMR's stated test is the ordinary one: non-public nature, commercial value, and confidentiality measures taken by the enterprise.
Primary sources: SAMR, six typical cases of trade secret infringement, 20 August 2026 (Chinese)
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
- Instrument (EN)
- Case three of six in 市场监管总局公布六起侵犯商业秘密典型案例 (SAMR publishes six typical cases of trade secret infringement). The case is headed 浙江省杭州市市场监管局查处孙某侵犯人工智能大模型商业秘密案
- Authority
- Order made by the Hangzhou Municipal Market Regulation Bureau (杭州市市场监管局), Zhejiang Province. Published nationally by the State Administration for Market Regulation (市场监管总局)
- Jurisdiction
- China
- Event type
- Administrative enforcement under existing unfair-competition law. Not a court judgment, not a statute, not an AI-specific regulation
- Date of legal event
- Order dated 28 May 2026. SAMR published the case on 20 August 2026 at 15:00, with 信息来源 given as 市场监管总局
- Legal basis cited
- Anti-Unfair Competition Law of the People's Republic of China (2019 revision), Article 9(1)(3) and Article 9(2). Penalty imposed under Article 21
- Penalty
- Order to stop the disclosure and authorised-use conduct, and a fine of RMB 350,000 (350000元)
- Materials at issue
- The rights holder's AI model exclusive prompt templates (专属提示词模板), review rules (审查规则) and annotation specifications (标注规范), which expert appraisal found combined into an integrated scenario intelligent review solution
- Separate proceedings
- The document states that the infringement by 法原（杭州）科技有限公司 is handled as a separate case (另案处理)
- Primary source
- https://www.samr.gov.cn/xw/zj/art/2026/art_45ceaace99a7454d97cd42f70ed404d5.html
What the regulator counted as the trade secret
The materials Sun sent were not source code. SAMR records them as the rights holder's AI model exclusive prompt templates, review rules and annotation specifications, and says that industry experts specially appraised the combination as forming an integrated scenario intelligent review solution.
That combination was found to possess the three ordinary statutory features: non-public nature, commercial value, and confidentiality management measures taken by the enterprise. On that basis it was treated as a new type of integrated technical trade secret in the AI vertical domain.
SAMR's own commentary is where the significance sits. It states that past AI intellectual-property enforcement was mostly confined to establishing rights in source code, that this case breaks the industry's code complex, and that it expressly finds natural-language integrated solutions and non-standard operational rules can each independently constitute trade secrets.
Read the scope carefully. SAMR describes this as the country's first AI vertical-domain large-model trade-secret case. That is narrower than a first AI trade-secret case of any kind, and much narrower than a first AI enforcement case. It is a claim about a category of model, made by the publishing regulator about its own case.
What happened
Sun joined a Hangzhou artificial intelligence company in July 2011 as a senior algorithm expert and left in July 2025. He led development of a vertical-domain AI intelligent review model and held the model's core confidential materials. He signed a specific confidentiality agreement on joining in July 2011, carrying obligations both during and after employment.
In December 2023, while still employed, he used his spouse's identity to register and in fact control 法原（杭州）科技有限公司. In June 2024 he sent the rights holder's prompt templates, review rules and annotation specifications to that company's research staff, for it to develop a comparable AI model.
On 28 May 2026 the Hangzhou Municipal Market Regulation Bureau ordered him to stop the disclosure and authorised-use conduct and fined him RMB 350,000. The document says the company's own infringement is dealt with separately, so nothing here states an outcome for it.
Why this matters for AI developers
The practical question this answers is which AI assets are protectable when the code is not the crown jewel. A great deal of applied AI value now sits in prompt design, review logic, evaluation criteria and labelling practice, none of which is source code and none of which is patentable in the ordinary course.
SAMR's framing sentence is worth noting on its own terms: it says enterprises increasingly tend to protect new types of technical information such as prompt engineering and Agent skill packages through trade secrets. That sentence describes what enterprises do, and sits in the case introduction as context. The materials actually at issue in this case were the prompt templates, review rules and annotation specifications. We are not asserting that Agent skill packages were adjudicated here, because the document does not say that.
The transferable point is the appraisal route. The secret was not any single document but the combination, assessed by industry experts as an integrated solution. Firms whose AI know-how is distributed across many small artefacts should note that the combination was what carried non-public nature and commercial value.
How this differs from SAMR's earlier AI cases
We already covered SAMR's February 2026 publication of five AI-related unfair-competition cases. This is a different publication, six months later, in a different area of the same statute: those were unfair-competition cases, this is a trade-secret batch, and this case turns on Article 9 rather than on the conduct provisions at issue there.
The two should be read together for direction of travel rather than merged. The February set showed the regulator applying general unfair-competition rules to AI conduct. This one shows it defining what counts as protectable AI subject matter in the first place.
What we did not verify
Opened and read in full: the SAMR page at samr.gov.cn, case three in its entirety, including the case introduction, the legal application and penalty paragraph, and the case commentary. The decode was corroborated as UTF-8 from the declared Content-Type, with no replacement characters, before any of it was read.
Not opened: the underlying Hangzhou Municipal Market Regulation Bureau decision itself, which SAMR summarises but does not reproduce; the expert appraisal; the confidentiality agreement; and the Anti-Unfair Competition Law text as distinct from the articles SAMR cites. The provisions are quoted as SAMR cites them, not as verified against the statute.
We refuse to claim: that this is a court judgment, that it creates any new obligation, that it is an AI regulation, or that it binds anyone beyond the party fined. We make no claim about the rights holder's or the respondent company's identity beyond what the document states, and we do not name the individual beyond the surname SAMR uses. We do not assert any outcome for 法原（杭州）科技有限公司, because the document expressly says that is a separate case.
Treat this as a scope decision, not a penalty story. A Chinese market regulator has said in terms that the protectable subject matter of an AI system is not limited to its code, and that prompt templates, review rules and annotation specifications can carry trade-secret status when they combine into something non-public, commercially valuable and actually kept confidential. If your AI confidentiality programme was written around a code repository, it is protecting the wrong assets.
Source File
https://www.samr.gov.cn/xw/zj/art/2026/art_45ceaace99a7454d97cd42f70ed404d5.html
Open the SAMR page and go to case three, headed 浙江省杭州市市场监管局查处孙某侵犯人工智能大模型商业秘密案. Confirm three things. First, the penalty paragraph gives the order date as 2026年5月28日 and the fine as 350000元, imposed under 第二十一条 of the 反不正当竞争法（2019年修正）, with the breach under 第九条第一款第（三）项 and 第九条第二款. Second, the case commentary contains 本案系全国首例人工智能垂类大模型商业秘密案件 and the phrase 代码情结. Third, the page header shows 发布时间：2026-08-20 15:00 and 信息来源：市场监管总局, which is the publication date rather than the date of the order.
expressly finding that natural-language integrated solutions and non-standard operational rules can each independently constitute trade secrets · SAMR case commentary, 20 August 2026, author's translation from the Chinese
FAQ
Is this a court judgment?
No. It is an administrative penalty decision by the Hangzhou Municipal Market Regulation Bureau, published nationally by SAMR as a typical case. SAMR publishes typical cases to guide enforcement practice; the document is a summary of the case, not a judgment and not a binding precedent in the common-law sense.
Does this create new obligations for AI companies?
No. The legal basis is the Anti-Unfair Competition Law as revised in 2019, and the elements applied are the ordinary trade-secret elements. Nothing in the document creates a new duty. What changed is the regulator's view of what subject matter those existing rules can cover.
Were Agent skill packages held to be trade secrets in this case?
Not on the text. The document mentions prompt engineering and Agent skill packages in an introductory sentence describing what enterprises increasingly protect through trade secrets. The materials actually at issue were the AI model's exclusive prompt templates, review rules and annotation specifications.
What does vertical-domain mean here?
The model in question is described as a vertical-domain AI intelligent review model, meaning a model built for a specific application area rather than a general-purpose one. SAMR's first-of-kind claim is framed around that category, and we have kept that qualifier rather than dropping it.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.