Why must a CRA threat model include AI attacks when the product has AI features?
The CRA covers the whole product with digital elements, whether a function is conventional code or an AI model. Once a product contains AI, the cybersecurity risk assessment required by Article 13(2) has to account for the new attack surface, and the Annex I Part I(2) requirements (protection against unauthorised access, integrity of data and commands, availability of essential functions, limited attack surfaces, recording and monitoring of internal activity) apply to the AI components as well.
- Legal link: CRA Article 12 provides that products also classified as high-risk AI systems under Article 6 of the AI Act are deemed to comply with the cybersecurity requirements of AI Act Article 15 where they meet the CRA Annex I essential requirements and this is shown in the EU declaration of conformity.
- Attacks named in AI Act Article 15(5): data poisoning, model poisoning, adversarial examples or model evasion, confidentiality attacks and model flaws. That list is the minimum the AI part of a CRA threat model should cover.
- CRA recital 51: when describing cyber resilience risks of high-risk AI systems, it names AI-specific vulnerabilities such as data poisoning and adversarial attacks.
- Timeline: CRA Article 14 reporting obligations have applied since 11 September 2026 and the remaining obligations apply from 11 December 2027. After Regulation (EU) 2026/1744, AI Act high-risk obligations apply from 2 December 2027 for Annex III uses and from 2 August 2028 for products regulated under Annex I.
So AI governance does not replace the CRA, and the CRA not mentioning AI by name does not exempt AI: a compromised AI component is a compromised product.
What AI-specific attack surface does a product with AI features have?
The table follows the attack classes in NIST AI 100-2 E2025 and the OWASP Top 10 for LLM Applications 2025, mapped to entry points that actually appear in products. Not every row applies to every product; assess each against the product architecture.
| AI attack | Entry point in the product | Possible consequence | Reference |
|---|---|---|---|
| Data poisoning, model poisoning | Training data sources, pre-trained models, model files shipped in OTA updates | Backdoored or degraded model | AI Act Art. 15(5); NIST poisoning; OWASP LLM04 |
| Adversarial examples / evasion | Sensor, image and voice input | Misclassification, safety functions defeated | AI Act Art. 15(5); NIST evasion |
| Direct and indirect prompt injection | Voice assistants, chat interfaces, features that read external documents or web pages | Restrictions bypassed, unauthorised actions triggered | NIST prompt injection; OWASP LLM01 |
| Excessive agency (model can call device functions) | AI agent that switches functions, changes settings or calls APIs | Injected instructions become device actions | OWASP LLM06 |
| Confidentiality attacks: model theft, membership inference, data extraction | Cloud inference APIs, model files on the device | Model or training data exposed | AI Act Art. 15(5); NIST privacy attacks |
| Unbounded consumption | Remote data processing (cloud AI service) | Outage, runaway cost | OWASP LLM10; Annex I Part I(2)(h) |
| AI supply chain | Third-party models, libraries, model hubs | Known weaknesses or malicious components imported | NIST GenAI supply chain attacks; OWASP LLM03 |
How do you build AI attacks into the CRA threat modelling process?
We recommend adding four steps to the existing CRA threat modelling process rather than starting a separate AI document set:
- List AI components as assets: model files and versions, training and fine-tuning data sources, system prompts, inference endpoints and AI suppliers. Put them in the component inventory too. Annex I Part II(1) requires an SBOM covering at least top-level dependencies; including models and AI libraries in the same inventory is our practical recommendation, not a format the regulation prescribes.
- Draw AI data flows and trust boundaries: which inputs come from users, which come from external documents or web pages (the source of indirect prompt injection), and whether model output drives device actions directly.
- Apply the AI attack list item by item: use the table above, rate likelihood and impact for each entry point and decide treatment; high-risk items need a technical control and a test.
- Connect AI weaknesses to vulnerability handling: jailbreaks or prompt-injection bypasses that can be exploited to affect product security go through PSIRT like any other vulnerability, including the decision whether it is an actively exploited vulnerability to be reported under Article 14.
The threat model must follow model versions: a new model, a new fine-tune or a new AI feature triggers reassessment.
Which technical controls apply, and what evidence should you keep?
The requirement column follows the official text; the controls are our practical recommendations.
| Requirement | Control for AI components | Evidence |
|---|---|---|
| Annex I Part I(2)(f): integrity of data, commands, programs and configuration | Signed model files verified at update, provenance and hashes of training data | Signature verification logs, data source register |
| Annex I Part I(2)(d): protection against unauthorised access | Authenticated and authorised inference APIs; least privilege and human confirmation for functions an AI agent can call | Permission design, test records |
| Annex I Part I(2)(h): availability of essential functions, resilience against denial of service | Inference rate limits, input length limits, abnormal usage alerts | Configuration baseline, alert records |
| Annex I Part I(2)(j): limit attack surfaces | Disable unneeded AI interfaces; filter and isolate external content before it reaches the model | Design review records |
| Annex I Part I(2)(l): record and monitor internal activity | Security logs of prompts, model responses and triggered actions (with personal data minimised) | Event logs, monitoring rules |
| Art. 13(2) risk assessment; AI Act Art. 15(5) | Adversarial testing, prompt-injection testing, model theft and extraction testing | Test plans and reports, threat model versions |
Who does what across R&D, security, product and AI suppliers, and where do Winners and Ahura fit?
- Product and R&D: attach data flows and trust boundaries to every AI feature request; a model change or new fine-tune triggers reassessment.
- Security and PSIRT: maintain the AI attack list, schedule adversarial and prompt-injection testing, and bring AI weaknesses into vulnerability handling and Article 14 reporting decisions.
- Procurement: agree vulnerability notification, data use and update responsibilities with AI model and cloud AI service suppliers (in line with the Article 13(5) due diligence on third-party components).
- Compliance: decide whether the product is also a high-risk AI system and which conformity route under Article 12 applies.
Winners Consulting handles scope and classification decisions, threat modelling workshops, control design and evidence planning; adversarial testing can be run by the client's team or an independent tester, with us setting the scope and reviewing results. Ahura (POC version) covers the CRA flow of threat modelling, SBOM vulnerability management, 24/72-hour reporting deadline tracking and technical documentation sign-off, with all data kept on the company's own premises.
Already working on the CRA: where do you start with AI threat modelling?
Start with three things:
- Inventory: list the AI components in each product (in-house models, third-party models, cloud AI services, AI agent features) and the actions they can take.
- Extend the threat model: run one round of threat modelling on products with AI using the attack list on this page, starting with the two high-impact cases where model output drives device actions and where model files ship with updates.
- Connect vulnerability handling: add an AI weakness category to the PSIRT process and agree a vulnerability notification contact with each AI supplier.
Once these are in place, add the tests as a fixed gate before release so the CRA technical documentation can cite the records directly.
AI attack defence for other services
Official sources
- Regulation (EU) 2024/2847 (Cyber Resilience Act) — EUR-Lex
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 15 — EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — 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
