Cyber Resilience Act and the AI Act: When Your Product Is Both

Cyber Resilience Act and the AI Act: high-risk AI as a product with digital elements under the EU Cyber Resilience Act

Cyber Resilience Act and the AI Act increasingly apply to the same product. A high-risk AI system under the AI Act (Regulation (EU) 2024/1689) is almost always a product with digital elements under the EU Cyber Resilience Act (Regulation (EU) 2024/2847) — and Article 12 of the CRA lets one set of cybersecurity evidence serve both. Here is how to avoid doubling the work.

TL;DR

  • A high-risk AI system is almost always a product with digital elements, so the Cyber Resilience Act and the AI Act both apply.
  • Article 12 lets one set of CRA cybersecurity evidence serve the AI Act’s cybersecurity arm — but not its accuracy or robustness.

A growing share of products with digital elements (PDEs) now embed AI — biometric access control, industrial vision, AI decision support. Many fall under both horizontal regimes at once. They were drafted for different objectives — the CRA secures products across their lifecycle; the AI Act regulates AI systems by risk — but they overlap when an AI system is also a PDE. The question is how to demonstrate compliance with both without doubling the work.

The dual scope of Cyber Resilience Act and the AI Act

A high-risk AI system under Article 6 of the AI Act is almost always a PDE under Article 3(1) of the CRA. The high-risk list — biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice — is virtually all software, often with hardware and remote data processing. Where the dual classification applies, the manufacturer must meet the CRA’s essential requirements (Annex I), the AI Act’s high-risk requirements (risk management, data governance, documentation, and accuracy/robustness/cybersecurity), and both conformity assessments. Without coordination that means two risk assessments and two documentation sets — but the CRA was drafted with this overlap in mind.

How the Cyber Resilience Act and the AI Act integrate: Article 12

First, high-risk AI systems that fall within the CRA must comply with the CRA’s essential cybersecurity requirements — no carve-out. Second, where they do, they are deemed to comply with the cybersecurity requirements of Article 15 of the AI Act. Article 12(1) makes this conditional on three cumulative requirements: (1) the product fulfils Annex I Part I; (2) the manufacturer’s processes fulfil Annex I Part II; and (3) the level of protection required under AI Act Article 15 is demonstrated in the CRA EU declaration of conformity.

This is the legal hook that lets a single set of cybersecurity evidence serve both regimes — but only “to the extent covered.” AI Act Article 15 also covers accuracy and robustness; the CRA covers cybersecurity. So the CRA declaration satisfies the AI Act’s cybersecurity arm, not its accuracy or robustness arms — Article 12(1) is explicit that the presumption is without prejudice to those.

AI-specific cybersecurity risks in your CRA risk assessment

Recital 51 requires the CRA risk assessment for a high-risk AI-PDE to take in AI-specific vulnerabilities — data poisoning (manipulating training data), adversarial attacks (inputs crafted to cause errors), and risks to fundamental rights — mirrored by AI Act Article 15(5). In practice, for a CRA risk assessment under Article 13(2): the threat model extends across the AI lifecycle (training-data integrity, model provenance, pre-trained components); adversarial robustness becomes a security control (input validation, anomaly detection, output-bounds checking, adversarial training) traceable to Annex I; logging gains drift and anomaly detection; and update mechanisms must cover model updates, re-training and prompt/system-message changes, not just code. For third-party AI components (foundation models, embeddings, off-the-shelf vision libraries), the component due-diligence duty under Article 13(5) applies. Ground the risk step in our CRA threat modeling guide.

The practical playbook for Cyber Resilience Act and the AI Act

  1. Map to both scopes deliberately — is it high-risk under AI Act Article 6, and a PDE under CRA Article 3(1) (and in which class)? Document both.
  2. Run one integrated risk assessment covering AI Act Article 9 risks (health, safety, fundamental rights), AI-specific cyber vulnerabilities, and all of CRA Annex I — one artefact mapped to both.
  3. Build one technical-documentation set cross-referencing AI Act Annex IV and CRA Annex VII; the SBOM serves both (and NIS2 — see our NIS2 and CRA guide).
  4. Integrate vulnerability handling for the full AI stack — code, models, training-data pipelines and third-party components.
  5. Coordinate incident reporting — CRA Article 14 to ENISA via the Single Reporting Platform (24h/72h/14 days, from 11 September 2026) and AI Act Article 73 serious-incident reporting to the market surveillance authority; an incident can trigger both.
  6. Plan to the agreed timeline — build to the CRA for cybersecurity, use the Article 12 presumption as the bridge to AI Act Article 15, and align to the AI Act’s deferred high-risk dates.

What it means for different manufacturers

Biometric, identity and access-control systems are commonly both Annex III high-risk AI and important PDEs — the integrated approach applies in full. Industrial AI and computer vision often sit alongside machinery under the Machinery Regulation; under the agreed Digital Omnibus, AI-related health-and-safety requirements are folded into the Machinery Regulation, so the practical effect is one product-safety conformity assessment (machinery, now incorporating AI) plus the CRA for cybersecurity. General-purpose AI model providers will face growing downstream demands for evidence of secure development and vulnerability handling (CRA Article 13(5)/(6)). Medical AI stays under the MDR — the CRA’s MDR carve-out means the CRA does not apply directly, but the AI Act still does.

Note on timing: under the Digital Omnibus agreement, the AI Act’s high-risk obligations were deferred — to around December 2027 for standalone Annex III systems and around August 2028 for embedded Annex I systems — while the CRA dates (reporting 11 September 2026, full application 11 December 2027) are unchanged. Confirm the final dates against the adopted text.

● Invest in CRA-grade cybersecurity now — the same evidence serves the AI Act’s cybersecurity arm.

Run one integrated risk assessment for both regimes in our CRA Compliance Software, try our free CRA online check tool, or book a demo.

Conclusion: Cyber Resilience Act and the AI Act together

The two regimes are not redundant — the AI Act addresses AI-specific risks (beyond cybersecurity), the CRA addresses product cybersecurity comprehensively. But where a high-risk AI system is also a PDE, the Cyber Resilience Act and the AI Act intersect through CRA Article 12 and AI Act Article 15, and an integrated approach is essential. Invest in CRA-grade cybersecurity now, knowing the same investment increasingly satisfies the AI Act’s cybersecurity dimension too.

Frequently asked questions

Does meeting the CRA satisfy the AI Act’s cybersecurity requirements?

Under CRA Article 12, a compliant CRA EU declaration of conformity gives a presumption of compliance with AI Act Article 15’s cybersecurity requirement — but not its accuracy or robustness requirements.

Which AI-specific risks must a CRA risk assessment cover?

Data poisoning, adversarial inputs, model evasion and confidentiality attacks, plus risks to fundamental rights, across the AI lifecycle including training data and third-party models.

When do the AI Act high-risk obligations apply?

Under the Digital Omnibus agreement they were deferred to around December 2027 (standalone Annex III) and around August 2028 (embedded Annex I); the CRA dates are unchanged. Confirm against the adopted text.

Related to this article

Check for free

Does the Cyber Resilience Act apply to you? →

Product Check and Readiness Check, no login.

Glossary

Cyber Resilience Act glossary →

30 key terms explained, each with its legal reference.

Where does your product stand on the Cyber Resilience Act?

A first answer in minutes — or see the software that turns the legal text into a workflow.

Free CRA online checkExplore the CRA software


Last updated on 23 October 2026 · Sources: Regulation (EU) 2024/2847 (CRA) — EUR-Lex, AI Act (EU) 2024/1689 — EUR-Lex, ENISA — Product security and certification

Alexander Aigner

About the author

Alexander Aigner

Product Security Expert, COMPLYD

Product security expert who has secured products in two of the most heavily regulated industries — medical and automotive. He helps teams meet the EU Cyber Resilience Act without slowing development.

LinkedIn profile