One product, two rulebooks: the cybersecurity of high-risk AI systems between article 15 AI act and the cyber resilience act after the digital omnibus

Abstract High-risk AI systems distributed as software or embedded in connected hardware are simultaneously subject to two horizontal security rulebooks: the accuracy, robustness and cybersecurity requirements of Art. 15 of the AI Act and the essential cybersecurity requirements of the Cyber Resilience Act (CRA). The Union legislature anticipated this overlap and built a bridge: under Art. 12 CRA—since July 2026 mirrored in Art. 42(3) AI Act, as inserted by the Digital Omnibus on AI (Regulation (EU) 2026/1744)—products with digital elements that are classified as high-risk AI systems are deemed to comply with the cybersecurity requirements of Art. 15 AI Act if they satisfy the CRA’s essential requirements and the achievement of the Art. 15 protection level is demonstrated in the CRA declaration of conformity, with the conformity assessment, as a rule, integrated into the AI Act procedure. The Omnibus has re-sequenced the two regimes: the high-risk obligations for stand-alone Annex III systems, deferred to 2 December 2027, now become applicable nine days before the CRA’s main obligations on 11 December 2027, whereas high-risk AI embedded in Annex I products follows on 2 August 2028, and the CRA’s reporting duties have bound manufacturers—including of AI products already on the market—since 11 September 2026. Drawing on the consolidated legislative texts, the Commission’s CRA guidance of July 2026 and the standardisation work of CEN-CENELEC JTC 13 and JTC 21, this article maps the scope overlap, analyses the mechanics and limits of the bridge—the accuracy and robustness carve-out, the narrowly drawn derogation for important and critical products in Art. 12(3) CRA, and the gap between AI-specific attack vectors and the CRA’s generic essentials—and argues that the bridge is coherent on a reading, anchored in the recitals of both acts, under which the adversarial-machine-learning threat model enters the CRA’s risk-based essentials. It translates the analysis into an integrated compliance architecture with a structured requirement-to-evidence mapping and formulates recommendations addressed separately to the Commission and the standardisation bodies, and to providers and manufacturers.