The EU AI Act, Regulation (EU) 2024/1689, sets a dedicated cybersecurity duty for AI systems the Regulation classifies as "high-risk" and a parallel one for the largest general-purpose AI (GPAI) models. Per the Regulation's own Article 15, high-risk AI systems must "achieve an appropriate level of accuracy, robustness, and cybersecurity" and stay resilient "throughout their lifecycle" against named AI-specific attacks — data poisoning, model poisoning, adversarial examples and confidentiality attacks. Providers of GPAI models with "systemic risk" get a separate, narrower duty under Article 55(1)(d): "ensure an adequate level of cybersecurity protection" for the model itself and its physical infrastructure. This guide covers both duties, who they apply to, when they bite after a mid-2026 amendment reshuffled the dates, and how two existing certification routes — Cyber Resilience Act compliance and EUCC certification — can carry over.
Key takeaways
- The EU AI Act, Regulation (EU) 2024/1689, sets a dedicated cybersecurity duty for AI systems the Regulation classifies as "high-risk" and a parallel one for the largest general-purpose AI (GPAI) models.
- Para. 1 sets the top-level standard: an "appropriate level" of accuracy, robustness and cybersecurity, maintained consistently "throughout their lifecycle" — not just at the point of CE marking.
- Two routes lead to high-risk status.
- Article 113, as originally enacted, set a general application date of 2 August 2026, with Article 6(1) obligations from 2 August 2027.
What does Article 15 require of high-risk AI systems?
Article 15 of the AI Act's Official Journal text is structured as a lifecycle obligation, not a one-time check:
- Para. 1 sets the top-level standard: an "appropriate level" of accuracy, robustness and cybersecurity, maintained consistently "throughout their lifecycle" — not just at the point of CE marking.
- Para. 2 asks the Commission to encourage benchmarks and measurement methods for accuracy and robustness.
- Para. 3 requires accuracy levels and the metrics used to measure them to be declared in the system's instructions for use.
- Para. 4 covers robustness against errors, faults and inconsistencies, allows "technical redundancy solutions, which may include backup or fail-safe plans", and — for systems that keep learning after deployment — requires measures against biased-output feedback loops.
- Para. 5 is the cybersecurity paragraph proper: high-risk systems must be "resilient against attempts by unauthorised third parties to alter their use, outputs or performance by exploiting system vulnerabilities." The technical solutions must be "appropriate to the relevant circumstances and the risks" and include, "where appropriate," measures against data poisoning, model poisoning, adversarial examples or model evasion, confidentiality attacks and model flaws.
Para. 5 names attacks on the AI itself. Ordinary device security — secure boot, update integrity, network hardening — still counts, because an attacker who owns the firmware also owns the model.
When is an AI function in a device "high-risk"?
Two routes lead to high-risk status. Article 6(2) and Annex III cover stand-alone use cases in eight areas: biometrics; critical-infrastructure safety components; education; employment and worker management; access to essential private and public services; law enforcement; migration, asylum and border control; and administration of justice and democratic processes. For device makers, point 2 matters most: "safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity."
Article 6(1) covers AI that is a safety component of a product, or is itself a product, under the EU product laws listed in Annex I — but only where that product "is required to undergo a third-party conformity assessment" under that law. Regulation (EU) 2026/1744, the Digital Omnibus on AI, changed this route in three ways that matter to hardware teams:
| Change (2026/1744) | Effect for device makers |
|---|---|
| New Art. 6(1a) and (1b) | AI used "solely" for user assistance, performance optimisation, service efficiency, automation, convenience or quality control is not a safety component — unless its "failure or malfunctioning … would endanger health and safety" |
| New Art. 6(1c) | A third-party assessment required "solely due to risks other than risks to health and safety", such as radio spectrum or electromagnetic interference, does not trigger Art. 6(1) |
| Annex I amended | The old Machinery Directive entry is deleted from Section A, and Machinery Regulation (EU) 2023/1230 is added to Section B |
Section B status matters because, under Article 2(2), only Article 6(1), Articles 102 to 109 and Article 112 of the AI Act apply to Section B products. For machinery, the AI requirements will arrive through the Machinery Regulation itself: the Omnibus tells the Commission to add them to its Annex III by delegated acts that "shall apply by 2 August 2028." See our Machinery Regulation guide. Medical devices, in vitro diagnostics, radio equipment, toys and lifts stay in Section A.
What does Article 55 require of systemic-risk GPAI models?
A second, narrower cybersecurity duty sits in Chapter V, aimed at providers of general-purpose AI models the Act treats as carrying "systemic risk" (broadly, the most capable foundation models, assessed partly against a compute threshold). Article 55(1) layers four extra obligations on top of the baseline GPAI duties in Articles 53-54: model evaluation including adversarial testing, Union-level systemic-risk assessment and mitigation, serious-incident reporting to the AI Office, and — in Article 55(1)(d) — a duty to "ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model." Article 55(2) lets providers rely on codes of practice until a harmonised standard is published. A device maker that only integrates a GPAI model downstream is not the provider of that model.
Which dates apply after the Digital Omnibus?
Article 113, as originally enacted, set a general application date of 2 August 2026, with Article 6(1) obligations from 2 August 2027. Regulation (EU) 2026/1744, adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, replaced Article 113(c). Its recital 40 gives the reason: "the delayed availability of standards, common specifications, and alternative guidance and the delayed establishment of national competent authorities." The Commission's AI Act policy page tracks the same timeline.
| Date | What applies |
|---|---|
| 1 August 2024 | AI Act enters into force (20 days after 12 July 2024 OJ publication) |
| 2 February 2025 | Chapters I-II: Article 5 prohibited practices (most) and AI-literacy duties |
| 2 August 2025 | Chapter III Section 4, Chapter V (incl. Article 55 GPAI cybersecurity), Chapter VII, Chapter XII, Article 78 |
| 27 July 2026 | Articles 102-110 (the AI Act's amendments to sectoral laws such as aviation security, vehicle type-approval and civil aviation rules) |
| 11 September 2026 | CRA Article 14 vulnerability and incident reporting |
| 2 December 2026 | The two Article 5 prohibited-practice points added by the Digital Omnibus |
| 11 December 2027 | CRA applies in full, including Annex I essential requirements |
| 2 December 2027 | Chapter III Sections 1-3 (incl. Article 15) for Annex III high-risk systems |
| 2 August 2028 | Chapter III Sections 1-3 (incl. Article 15) for Annex I high-risk (product-embedded) systems |
The CRA dates come from its Article 71. For a product that is both a CRA product and an Annex III high-risk AI system, the CRA's Annex I deadline (11 December 2027) arrives nine days after Article 15 does (2 December 2027), so one engineering program can serve both.
Which certification routes can carry over under Article 42?
A provider does not necessarily have to build an Article 15 cybersecurity case from scratch. Article 42 gives two cybersecurity routes (its paragraph 1 is a separate presumption about training data):
Article 42(2) — Cybersecurity Act certificationA high-risk AI system certified, or covered by a statement of conformity, under an EU cybersecurity certification scheme adopted under Regulation (EU) 2019/881 (the Cybersecurity Act) — such as the EUCC scheme — is presumed to meet Article 15's cybersecurity requirements, "in so far as" the certificate or statement covers them. See our EUCC / Common Criteria guide. Article 42(3) with CRA Article 12 — Cyber Resilience Act complianceWhere a product with digital elements is both in scope of the Cyber Resilience Act (Regulation (EU) 2024/2847) and classified as a high-risk AI system, CRA Article 12(1) deems it compliant with AI Act Article 15's cybersecurity requirements once the manufacturer (a) meets the CRA's Annex I Part I essential cybersecurity requirements, (b) runs the Annex I Part II vulnerability-handling processes, and (c) demonstrates that in the CRA's EU declaration of conformity. The Digital Omnibus copied that rule into the AI Act as a new Article 42(3), so both texts now say it. CRA Article 12(2) then applies the AI Act's own Article 43 conformity-assessment procedure. CRA Article 12(3) carves out an exception: "important" (Annex III) and "critical" (Annex IV) CRA products keep the CRA's stricter conformity-assessment route. See our CRA Annex I guide for the baseline this depends on.
One limit is easy to miss. CRA Article 12(1) opens "Without prejudice to the requirements relating to accuracy and robustness set out in Article 15." The bridge covers cybersecurity only; accuracy metrics, robustness testing and feedback-loop controls still need their own evidence under the AI Act. For a device or firmware maker already building to CRA Annex I, the cybersecurity work largely overlaps, and the CRA declaration of conformity — not a separate AI-specific cybersecurity dossier — is what Article 12 asks for. See our Cyber Resilience Act guide.
Which device controls answer Article 15(5)'s named attacks?
The AI Act names the attacks, not the controls. The table below is this guide's mapping for an AI model shipped inside device firmware; harmonised standards, once cited in the Official Journal, will set the presumption of conformity.
| Article 15(5) attack | Typical device-side control | Evidence to keep |
|---|---|---|
| Data poisoning | Provenance and integrity checks on training and fine-tuning data; access control on data pipelines | Dataset lineage records |
| Model poisoning (pre-trained components) | Pin and hash third-party models and weights; review their source | An AI-BOM listing models, versions and hashes |
| Adversarial examples / model evasion | Input range checks, sensor plausibility checks, adversarial robustness tests, safe fallback on low confidence | Test reports per release |
| Confidentiality attacks | Encrypted model storage, secure boot, rate limits on inference APIs | Threat model, penetration test results |
| Model flaws | Regression tests on accuracy metrics, monitoring after deployment | Accuracy figures declared under para. 3 |
Classic firmware vulnerabilities sit under all five rows. An exploitable bug in the inference service's HTTP server or TLS library lets an attacker swap the model outright, so the firmware SBOM and its CVE tracking belong in the Article 15 file too. ENISA's multilayer framework for AI cybersecurity gives a broader control catalogue.
What does a worked example look like?
Synthetic example, for illustration only. The company, product and figures are invented.
A maker sells a pump-station controller with an AI model that detects pressure anomalies and trips a safety valve for a water utility.
ClassificationThe model is a safety component in the supply of water, so Annex III point 2 applies, and Article 15 applies from 2 December 2027. The controller is also a CRA product with digital elements. CRA routeThe team meets CRA Annex I Part I and Part II and states Article 15 cybersecurity in the CRA declaration of conformity, using AI Act Article 43 for the assessment. What the bridge does not coverThe team declares the model's detection rate and false-trip rate in the instructions for use (Article 15(3)) and documents the fail-safe that closes the valve on sensor disagreement (Article 15(4)). Article 15(5) controlsModel weights are signed and checked at boot; the AI-BOM lists the open-source base model and its hash; adversarial tests on spoofed sensor inputs are run for each release. Vulnerability handlingThe firmware SBOM is matched against new CVEs; a KEV-listed flaw in the web server triggers an update under the CRA Part II process.
What are the common pitfalls?
- Using 2 August 2026 as the deadline. It is no longer the Article 15 date for any high-risk category.
- Assuming the CRA covers all of Article 15. It covers cybersecurity only, "without prejudice" to accuracy and robustness.
- Classifying every AI feature as high-risk, or none. Article 6(1a) excludes convenience and quality-control functions unless their failure endangers health or safety.
- Missing the machinery move. AI in machinery now follows the Machinery Regulation's delegated acts, not AI Act Chapter III directly.
- Testing the model and not the firmware around it. A model is only as protected as the boot chain and update path that load it.
Where should a team start?
Classify firstDetermine whether your system is in scope of Annex III (stand-alone high-risk use case) or Article 6(1)/Annex I (embedded in an already-regulated product), and check the new Article 6(1a)–(1c) exclusions — the compliance clock differs by eight months between the two routes. Treat Article 15(5)'s named attacks as a checklist, not an afterthoughtData poisoning, model poisoning, adversarial examples and confidentiality attacks are the specific technical scope the article names; general IT security controls alone will not demonstrate compliance. Check for an existing bridge before building new paperworkIf the product is also a CRA-scope product with digital elements, or already holds an EUCC-type certificate, map what Article 42 and CRA Article 12 let you reuse before starting an AI Act-specific cybersecurity assessment. Track the amended dates, not the original ones2 December 2027 (Annex III) and 2 August 2028 (Annex I) are the Article 15 dates following Regulation (EU) 2026/1744. If you provide a GPAI model, don't wait for the high-risk datesArticle 55's cybersecurity duty for systemic-risk models has applied since 2 August 2025 regardless of the Digital Omnibus delay to Chapter III.
SoftWSP covers the firmware part of that file. Its device and firmware review extracts an SBOM from the firmware image as CycloneDX 1.6 JSON. It matches the components against NVD, the CISA KEV catalog, FIRST EPSS and CISA CSAF ICS advisories, and drafts a cited report that an engineer reviews. That is input for CRA Annex I Part II vulnerability handling. It does not test AI models for poisoning or adversarial robustness, and it does not write the declaration of conformity: see Device & Firmware Security.
This guide summarises Regulation (EU) 2024/1689, its amending Regulation (EU) 2026/1744 and Regulation (EU) 2024/2847 as published in the Official Journal, read from the EU Publications Office, together with the European Commission's AI Act policy page. The control mapping is this guide's own engineering reading, not legal text, and nothing here is legal advice.
Put this guide to work
Frequently asked questions
What does the EU AI Act actually require for cybersecurity?
Article 15 of Regulation (EU) 2024/1689 requires that high-risk AI systems 'be designed and developed in such a way that they achieve an appropriate level of accuracy, robustness, and cybersecurity, and that they perform consistently in those respects throughout their lifecycle.' Per the Regulation's own text, Article 15(5) adds that these systems must be 'resilient against attempts by unauthorised third parties to alter their use, outputs or performance,' naming AI-specific attack classes the technical measures must address: data poisoning, model poisoning, adversarial examples, and confidentiality attacks. Separately, Article 55(1)(d) puts a parallel cybersecurity duty on providers of general-purpose AI (GPAI) models that carry 'systemic risk,' requiring them to 'ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model.'
Which AI systems count as 'high-risk' and have to meet Article 15?
Article 6 and Annex III classify a system as high-risk. Annex III lists eight areas: biometric identification/categorisation; safety components in critical infrastructure (digital infrastructure, road traffic, water/gas/heat/electricity supply); education and vocational training; employment, worker management and self-employment access; access to essential private and public services (credit scoring, insurance pricing, public-benefit eligibility, emergency dispatch); law enforcement; migration, asylum and border control; and administration of justice and democratic processes. Separately, Article 6(1) and Annex I cover AI that is a safety component of a product already regulated under EU product-safety law (machinery, medical devices, toys, lifts and similar) — these follow a different compliance timeline from the Annex III list.
What is the real compliance deadline for high-risk systems — 2026, 2027 or 2028?
The Regulation's original Article 113 text set 2 August 2026 as the general application date for Chapter III's high-risk rules. That date was pushed back by a later amending regulation, Regulation (EU) 2026/1744 (the 'Digital Omnibus on AI'), published in the Official Journal on 24 July 2026 and in force since 27 July 2026. As amended, Article 113(c) now applies the Chapter III high-risk requirements — including Article 15 — from **2 December 2027** for Annex III (stand-alone) high-risk systems, and from **2 August 2028** for Annex I (product-embedded) high-risk systems. What did not move: the Article 5 prohibited-practice rules (2 February 2025, plus two practices added by the Digital Omnibus phasing in 2 December 2026) and the GPAI provider obligations in Chapter V, including Article 55's cybersecurity duty (2 August 2025).
If I already comply with the Cyber Resilience Act or hold an EUCC certificate, do I have to redo cybersecurity work for the AI Act?
Not necessarily. Article 42 of the AI Act sets up two presumption-of-conformity routes for Article 15's cybersecurity requirement. Article 42(2): a high-risk AI system certified, or covered by a statement of conformity, under an EU cybersecurity certification scheme adopted under Regulation (EU) 2019/881 (the EU Cybersecurity Act — see our [EUCC guide](/guides/eucc-common-criteria-certification/)) is presumed to meet Article 15's cybersecurity requirements to the extent the certificate covers them. Article 42(3): where a product with digital elements is in scope of the Cyber Resilience Act (Regulation (EU) 2024/2847) and is also a high-risk AI system, Article 12 of the CRA does the bridging, and Regulation (EU) 2026/1744 copied the same rule into the AI Act as this new Article 42(3) — CRA Article 12 deems the product compliant with AI Act Article 15 once the manufacturer meets the CRA's Annex I Part I essential requirements, follows Annex I Part II vulnerability-handling processes, and records that in the CRA's EU declaration of conformity; the AI Act's own Article 43 conformity-assessment procedure then applies in place of the CRA's. See our [Cyber Resilience Act guide](/guides/cyber-resilience-act/) for the Annex I baseline this bridge relies on. One limit: CRA Annex III/IV 'important' and 'critical' products keep their stricter CRA-specific conformity assessment even when this bridge otherwise applies.
Sources
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act) — Official Journal text
- EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI, amending the AI Act's application dates) — Official Journal text
- European Commission: AI Act — regulatory framework and application timeline
- EU Artificial Intelligence Act (artificialintelligenceact.eu) — consolidated article-by-article text, incl. Articles 15, 42, 55 and 113
- Cyber Resilience Act, Article 12 — full text (craact.eu mirror)
- Gibson Dunn: EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes
- EUR-Lex: Regulation (EU) 2024/2847 (Cyber Resilience Act) — Article 12 (high-risk AI systems) and Article 71 (application dates)
About this guide
SoftWSP Regulation Desk
Covers the Cyber Resilience Act cluster: 34 guides.

In this cluster
Questions about this page?
