Why do AI channels affect whether a trade secret is legally protected?
Article 2 of Taiwan's Trade Secrets Act protects information only if it meets three conditions together: it is not known to persons generally involved with that type of information, it has actual or potential economic value because of its secrecy, and its owner has taken reasonable measures to keep it secret. Article 2(1)(c) of the EU Trade Secrets Directive (Directive (EU) 2016/943) likewise requires reasonable steps under the circumstances to keep the information secret.
In Taiwanese practice, Supreme Court civil judgment 102 Tai-Shang-Zi No. 235 read reasonable secrecy measures as classifying information by business need and letting only people at the right authorisation level know it, using means ordinarily available given the owner's resources. AI tools open new gaps in that classification:
- When staff paste confidential information into an external AI, the data leaves the authorised circle.
- An internal AI assistant that can retrieve every document effectively gives every user the highest access level.
If AI channels are left completely uncontrolled, the other side in a dispute may argue that the owner did not take reasonable secrecy measures. That does not guarantee losing the case, but it adds a weakness in proving it.
Through which AI attacks or uses can trade secrets leak?
| Leak path | Entry point | What may leak | Reference |
|---|---|---|---|
| Staff input into external AI | Generative AI on personal accounts, translation and summary extensions | Process parameters, formulas, quotes, customer lists | OWASP LLM02 Sensitive Information Disclosure |
| Code and document assistants | AI plug-ins in developer tools, AI features in cloud documents | Source code, design documents | OWASP LLM02, LLM03 Supply Chain |
| Direct and indirect prompt injection | Internal AI assistants that read mail, web pages or external documents | R&D documents the assistant can retrieve | NIST prompt injection; OWASP LLM01 |
| Retrieval permission bypass | R&D knowledge bases, vector databases | Confidential material the user is not entitled to see | OWASP LLM08 |
| Excessive agency of AI agents | Agents that read and write code repositories and file servers or send external mail | Bulk copying or sending of documents | OWASP LLM06 |
| Model extraction and copying | Inference APIs of proprietary models | Model behaviour and core algorithms | NIST privacy attacks (model extraction) |
| System prompt leakage | AI features in products | Proprietary rules and processes written into prompts | OWASP LLM07 |
| Insiders using AI to gather faster | AI used to summarise and compile large volumes of confidential files before leaving | Cross-department compilations of secrets | NIST misuse |
What controls bring AI into reasonable secrecy measures?
These controls are our practical recommendations. Their aim is to extend classification and need-to-know access to AI channels and to leave evidence for disputes.
| Aim | Control | Evidence |
|---|---|---|
| Limit AI channels by confidentiality level | Add an "allowed AI services" column to the classification scheme; top-secret material only into on-premises or approved enterprise models | Classification scheme, AI use policy |
| Prevent input leakage | AI gateway detects confidentiality markings and keywords and blocks them; block unapproved AI services and extensions | Gateway rules, block logs |
| Retrieval by original permissions | RAG retrieval applies each document's original access rights; top-secret material kept out of shared knowledge bases | Permission design, test records |
| Restrict AI agents | Agents get only the minimum permissions a task needs; bulk downloads and external sending need human confirmation | Permission list, confirmation records |
| Protect proprietary models | Model weights and training data listed as trade secrets; inference APIs rate-limited and monitored for abnormal queries | Asset list, monitoring records |
| Make the rules known | AI use rules in confidentiality agreements and work rules; training attendance recorded | Signed agreements, training records |
| Be able to prove it | AI usage logs retained and protected from tampering, ready for a dispute | Log retention policy |
How should legal, R&D, IT and HR divide the work?
- Legal and the trade secret office: maintain the classification scheme and AI use policy and handle evidence in disputes.
- R&D leads: decide which R&D material is top secret and may not enter any external AI.
- IT and security: build the AI gateway, RAG permission filtering, agent permissions and logging.
- HR: put AI use rules into onboarding, confidentiality agreements and exit procedures; check for abnormal AI use and downloads before departure.
- Procurement: agree with AI providers that data is not used for training, where it is stored and how it is deleted.
Which part do Winners take?
Winners Consulting inventories and classifies trade secrets, designs how AI channels become part of reasonable secrecy measures, writes the AI use policy and evidence plan, and integrates them with existing ISO 27001, ISO 27701 or TIPS systems instead of adding a separate document set. Technical controls such as the gateway and permission filtering are built by the company's IT team or its suppliers; we set the requirements and acceptance criteria. Ahura (POC version) is fully on-premises, keeping policy approvals and evidence records in the company's own environment.
Already running a trade secret programme: where do you start with AI?
- Inventory: the AI services and extensions staff actually use, and the internal AI assistants already live with the data they can retrieve.
- Handle the top secrets first: state that top-level process data, formulas and source code may not enter any external AI, and enforce it at the gateway.
- Update the paperwork: add AI use rules to confidentiality agreements, work rules and training so reasonable secrecy measures exist both in writing and in practice.
Then move on to internal AI assistant permissions and protection of proprietary models.
AI attack defence for other services
Official sources
- Trade Secrets Act — Laws & Regulations Database of the R.O.C. (Taiwan) (in Chinese)
- Directive (EU) 2016/943 on the protection of undisclosed know-how and business information (trade secrets) — EUR-Lex
- NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- OWASP Top 10 for LLM Applications 2025 — OWASP GenAI Security Project
