TISAX covers the organisation, so where does threat modelling for AI attacks on vehicle products belong?
Many automotive suppliers treat TISAX as the place where all security lives, but TISAX assesses how a company protects OEM information, prototypes and data, not the security engineering of the product itself. AI attacks therefore have to be handled on two layers:
- Product layer: AI functions in ECUs, cockpit systems, ADAS perception and connected-vehicle back ends are modelled through the ISO/SAE 21434 threat analysis and risk assessment (TARA, the methods in clause 15). Vehicle manufacturers need a UN R155 cyber security management system (CSMS) and type approval and usually extend TARA requirements to suppliers.
- Organisation layer (TISAX): engineers using generative AI tools on OEM drawings, specifications, prototype photos and test data, or training models on fleet data, are handling OEM confidential information and fall within the TISAX assessment scope.
- Functional safety: ISO/PAS 8800:2024 addresses safety risks from output insufficiencies and systematic errors of in-vehicle AI elements. It complements the 21434 security analysis and does not replace it.
Use one asset inventory and one threat vocabulary (the NIST AI 100-2 attack classes) for both layers so the story holds together in audits.
Which AI attacks should go into the TARA for a vehicle product?
UN R155 Annex 5 lists the threats and mitigations a vehicle-type risk assessment must at least consider; it is a minimum list, and threats specific to AI elements have to be added in the TARA. The table shows what we recommend assessing as a minimum:
| AI attack | Entry point in the vehicle | Possible consequence | Reference |
|---|---|---|---|
| Adversarial examples / evasion | Camera, radar and lidar perception input; deliberately altered road signs | Wrong object recognition, wrong ADAS decisions | NIST evasion attacks |
| Model poisoning, data poisoning | Fleet data uploads, labelling suppliers, pre-trained models | Backdoored model or failure in specific situations | NIST poisoning; OWASP LLM04 |
| Model file tampering | Over-the-air updates, diagnostic interfaces, flashing tools | Malicious or outdated model installed | R155 Annex 5 (update procedure threats); UN R156 software update management |
| Prompt injection | Cockpit voice assistants, in-vehicle AI that reads messages or web content | Unauthorised vehicle functions triggered, owner data exposed | NIST prompt injection; OWASP LLM01, LLM06 |
| Model theft, training data extraction | Model files in the vehicle, cloud inference APIs | Supplier's core model or owner data exposed | NIST privacy attacks; OWASP LLM02 |
| Back-end resource abuse | Connected-vehicle cloud AI services | Outage, runaway cost | OWASP LLM10 |
How do AI threats fit into the ISO/SAE 21434 TARA?
No separate AI process is needed. Add four things to the existing TARA:
- Asset identification: list model files and versions, training data sources, labelling suppliers, inference endpoints and prompts as assets, with their confidentiality, integrity and availability needs.
- Threat scenarios: apply the attack list above to each AI asset and write scenarios of the form: attacker, interface, action, damage.
- Attack feasibility and impact rating: rate with the 21434 method; judge the impact of adversarial examples and perception failure together with the functional safety team (ISO 26262, ISO/PAS 8800).
- Treatment and verification: map each scenario that needs treatment to a cybersecurity goal and a control and schedule tests; redo the TARA when the model is updated, retrained or sourced from a new supplier.
The TARA documents and CSMS evidence an OEM asks for can then cite these records directly.
On the TISAX layer, which controls and evidence are needed before OEM data enters an AI tool?
Requirements are those in the TISAX documents published by ENX and the VDA ISA. The table shows how we place AI use cases inside existing information security requirements; it does not quote ISA control numbers.
| Situation | Control | Evidence |
|---|---|---|
| Engineers paste OEM drawings or specifications into an external generative AI | Define by classification which levels may go into which AI services; confidential and prototype data only into approved enterprise or on-premises models | AI use policy, approved service list, gateway block logs |
| The AI service is an external IT service | Assess the supplier before use: training on customer data, storage location, retention, incident notification | Supplier assessment, contract clauses |
| Training models on fleet or test data | Access control on the training pipeline, records of data source and purpose; check the OEM contract allows it | Data flow diagram, authorisation records |
| Prototype photos and test images | Extend prototype protection to AI labelling and image tools; no uploads to unapproved platforms | Labelling supplier contracts, access logs |
| Accounts and usage | Personal accounts, disabled on leaving, abnormal usage alerts | Account list, alert records |
Who does what across product security, information security, R&D and suppliers, and which part do Winners take?
- Product security (CSMS) team: maintain the AI attack list, bring AI threats into the TARA and answer the OEM's cybersecurity interface agreement requirements.
- Information security (TISAX) team: set the AI use policy and approved list and handle access control and records for AI tools and OEM data.
- R&D and AI team: provide model, data source and update process information; a model release triggers reassessment.
- Procurement: agree data use limits and incident notification with AI service, labelling and model suppliers.
Winners Consulting determines TISAX scope and assessment level, brings AI use cases into the information security system, and helps product teams set up a TARA method that includes AI threats, with evidence planning. We advise only and do not perform TISAX assessments; assessments are carried out by audit providers recognised by ENX.
Already holding a TISAX label: where do you start with AI threat modelling?
- Build two inventories: the AI components in your products, and the AI tools staff actually use (including personal accounts).
- Control the data exits first: decide which AI services may receive OEM confidential and prototype data, and block unapproved services.
- Extend the TARA: pick one product with AI features, run a TARA round with the attack list on this page, turn it into a template and roll it out to other products.
These records are then ready for the next TISAX assessment or OEM audit.
AI attack defence for other services
Official sources
- ISO/SAE 21434:2021 Road vehicles — Cybersecurity engineering — ISO
- UN Regulation No. 155 — Cyber security and cyber security management system [2025/5] — EUR-Lex
- ISO/PAS 8800:2024 Road vehicles — Safety and artificial intelligence — ISO
- ENX Association — TISAX
- 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
