INSIGHT // 67 Comparative

Two Regimes, One Device: Swiss Medical Device Cybersecurity Where the AI Act Meets the MDR

Abstract: An AI-enabled medical device sold from Switzerland into the EU answers to two cybersecurity regimes that measure different things. The MDR secures the device in its IT environment, while the AI Act secures the model against manipulation of its own data. No harmonized standard yet translates either duty into testable requirements, and Switzerland adds a third layer with a 24-hour reporting clock of its own.
Plain Language Summary

This article examines how three sets of rules treat the cybersecurity of a medical device that contains artificial intelligence. The EU's Medical Device Regulation (MDR) sets safety and security requirements for the device. Where such a device needs a notified body, the EU AI Act applies on top of the MDR and sets separate requirements aimed at the machine-learning model itself. Switzerland is not an EU member. It applies the MDR's device requirements through its own ordinance (the Medizinprodukteverordnung, MepV) and has no AI Act, while its information-security statute (the Informationssicherheitsgesetz, ISG) imposes a cyberattack-reporting duty that can reach device manufacturers. The article describes where these regimes overlap, where they leave gaps, and why compliance with one does not establish compliance with the others.

Table of Contents
  1. The Two Compliance Perimeters and Where They Fail to Meet
  2. Where the AI Act's Security Duty Bites Harder Than the MDR's
  3. Switzerland's Position: MepV, ISG and the Absent AI Act
  4. Conformity Assessment Under Two Regimes at Once
  5. Strategic Considerations

US developers of AI-enabled devices arrive in Europe with a cybersecurity file already built. That file rests on section 524B of the FD&C Act for cyber devices, on the FDA's premarket cybersecurity guidance in its February 2026 revision and, for the model itself, on the threat vocabulary of the FDA's January 2025 draft guidance on AI-enabled device software functions, with its data poisoning, model inversion and model evasion.1FD&C Act § 524B; FDA cybersecurity guidance (3 February 2026); FDA draft AI guidance (7 January 2025). The instinct that travels with that file is that Europe asks the same question in a different dialect. It does not. The EU asks it twice, in two instruments that protect different objects with different tools. Switzerland, which sits outside the EU regulatory perimeter, asks it a third time and attaches a reporting clock of its own. A file that satisfies one of those questions can fail the other two without a single fact about the device changing.

1. The Two Compliance Perimeters and Where They Fail to Meet

The MDR's cybersecurity duty is short, and it is not addressed to the model. Section 17.2 of Annex I MDR requires that software be developed and manufactured in accordance with the state of the art, taking into account the principles of development life cycle, risk management, including information security, verification and validation. Section 17.4 of Annex I MDR requires the manufacturer to set out minimum requirements concerning hardware, IT networks characteristics and IT security measures, including protection against unauthorized access, necessary to run the software as intended. Section 23.4(ab) of Annex I MDR carries those minimum requirements into the instructions for use.2Regulation (EU) 2017/745 (MDR), Annex I, sections 17.1 to 17.4 and 23.4(ab); Art. 8; Art. 113. The object of protection is the device and its performance, and the perimeter is the device in its IT environment. The manufacturer's task is to specify what that environment must look like and to build a product that does not depend on it. MDCG 2019-16 Rev.1, the Medical Device Coordination Group's cybersecurity guidance of July 2020, follows that shape. Its elements are defense in depth, a security risk process running alongside safety risk management, and a joint responsibility, the guidance's own preferred term, shared among manufacturer, integrator, operator and user, with the operator expected to keep the environment secure and to install the patches the manufacturer prescribes.3MDCG 2019-16 Rev.1 (July 2020), section 2.6 on joint responsibility. It is a document about networks, access and updates. It was written four years before the AI Act existed, for a device whose logic is fixed at release.

What the MDR does not supply is a yardstick. Art. 8 MDR gives a presumption of conformity to devices that follow harmonized standards cited in the Official Journal. The standard the market treats as the reference for health-software security, EN IEC 81001-5-1:2022, sits in the Commission's standardization request to CEN and CENELEC with a deadline for adoption of 27 May 2028. As of publication its reference had not been published under the MDR. The ten amending decisions to the MDR harmonized-standards list through June 2026 cover gloves, sterilization, biological evaluation, risk management, implants and connectors, and none adds the security life-cycle standard.4Standardisation request M/575; Implementing Decision (EU) 2021/1182 and its amendments, none citing EN IEC 81001-5-1:2022. A manufacturer that follows the standard earns no presumption. A notified body that expects it is applying the state of the art rather than a cited standard, which is a different argument when a certificate is refused.

The AI Act's cybersecurity duty is addressed to something the MDR does not see. For an AI system that is a medical device, or a safety component of one, whose device requires notified-body assessment, Art. 6(1) and Annex I AI Act make the system high-risk. Art. 15(1) AI Act then requires that it achieve an appropriate level of accuracy, robustness and cybersecurity and that it perform consistently in those respects throughout its lifecycle. Art. 15(5) AI Act spells out what that means. The system must be resilient against attempts by unauthorized third parties to alter its use, outputs or performance by exploiting system vulnerabilities, with technical solutions that include, where appropriate, measures to prevent, detect, respond to, resolve and control for attacks that manipulate the training data set (data poisoning) or pre-trained components used in training (model poisoning), inputs designed to cause the model to make a mistake (adversarial examples or model evasion), confidentiality attacks or model flaws.5Regulation (EU) 2024/1689 (AI Act), Art. 6(1), Art. 9(2), Art. 13(3), Art. 15, Annex IV. The perimeter has moved. A poisoned training set alters a device's output without anyone gaining unauthorized access to the device. To the network and the access-control layer the MDR file describes, an adversarial input is an ordinary input. The attack surface the AI Act names is the model and the pipeline that produced it. The interests it protects under Art. 9(2)(a) AI Act are health, safety and fundamental rights, and no MDR risk file was ever asked to weigh the last of these.

The two perimeters could have been joined by a third instrument, and for medical devices they are not. Under Art. 12(1) of the Cyber Resilience Act (CRA), a high-risk AI system that meets the CRA's essential cybersecurity requirements, for the product and for the manufacturer's processes, and says so in its CRA declaration of conformity, is deemed to comply with the cybersecurity requirements of Art. 15 AI Act. The provision is a bridge, built so that product manufacturers would not have to derive the AI Act's security standard from first principles. Art. 2(2)(a) CRA then excludes from the CRA every product to which the MDR applies.6Regulation (EU) 2024/2847 (CRA), Art. 2(2)(a) and Art. 12(1). The joint FAQ of the MDCG and the European Artificial Intelligence Board, issued in June 2025, records the consequence in a single note. Medical devices are out of the CRA's scope.7AIB 2025-1 / MDCG 2025-6 (19 June 2025), Q22; not legally binding. For an AI-enabled medical device, the only EU instrument that translates Art. 15 AI Act into product-level cybersecurity requirements is closed, and the MDR text says nothing about poisoned data or evasion. The gap between the two regimes sits exactly where the model does.

The regime that names the attacks on a learning model is the one whose product-level cybersecurity bridge is closed to medical devices, while the regime that governs the device says nothing about the model.

The joint FAQ does what guidance can. It reads Art. 15 AI Act into the device file and expects manufacturers of high-risk medical AI to secure AI-specific assets such as training data sets and the trained model, to prevent data and model poisoning, and to treat cybersecurity as part of the risk management system and the quality management system, and therefore as subject to conformity assessment.7 That reading is not binding, as the document itself says, and it does not create the standard it presupposes. Whether a penetration test and a patch policy count as evidence of resilience against model evasion is left to the manufacturer's own risk assessment, which is to say it is left open.

2. Where the AI Act's Security Duty Bites Harder Than the MDR's

The FDA's cyber-device concept is gated on connectivity. Section 524B(c) of the FD&C Act reaches a device that includes software, has the ability to connect to the internet and carries characteristics that could be vulnerable to cybersecurity threats.1 Art. 15(5) AI Act has no such gate. It attaches to the high-risk AI system as such. A diagnostic model that never leaves an isolated workstation must still be resilient against manipulation of its training data, because the attack the provision names happens before the device is switched on. The FDA's January 2025 draft lists data poisoning, model inversion, model evasion and performance drift among the threats a submission should address. The AI Act writes a comparable list into the statute, then qualifies it twice, with "where appropriate" and with a requirement that the technical solutions be appropriate to the circumstances and the risks.5 The qualifiers do not soften the duty. They relocate the judgment about what the duty requires from the text to whoever assesses conformity.

The duty also runs longer than the MDR's. Art. 15(1) AI Act asks for consistent performance throughout the lifecycle. Art. 15(4) AI Act requires a system that continues to learn after being placed on the market to be built so as to eliminate or reduce as far as possible the risk of possibly biased outputs feeding back into future operations. Art. 9(2)(c) AI Act folds post-market monitoring data back into the risk process, and Annex IV AI Act requires the technical documentation to describe the cybersecurity measures put in place, next to the validation and testing procedures and the metrics used to measure accuracy and robustness.5 Art. 13(3)(b)(iii) AI Act then requires the instructions to deployers to state any known or foreseeable circumstance, under intended use or reasonably foreseeable misuse, that may lead to risks to health, safety or fundamental rights. Where section 23.4(ab) of Annex I MDR asks for minimum IT requirements, the AI Act's instructions ask what could be done to the model and how the deployer would notice.

None of this comes with a harmonized standard. The Commission's standardization request of 23 June 2025 asked CEN and CENELEC to deliver harmonized standards on cybersecurity specifications for AI systems, among the other deliverables it requested, by 31 August 2025. As of publication no harmonized standard supporting Art. 15 AI Act had been cited in the Official Journal. The Commission's own account expected the first deliverables from the standardization bodies in the course of 2026, with citation to follow a substantive review.8Commission Implementing Decision C(2025) 3871 (23 June 2025); no harmonized standard cited as of publication. Art. 17(1)(e) AI Act anticipates the gap. Where harmonized standards are not applied in full or do not cover a requirement, the quality management system must record the means used to ensure compliance. For cybersecurity that means every provider of a high-risk medical AI must construct its own account of what resilience against model evasion looks like and defend it before a body that has no cited standard to measure it against either. Common specifications under Art. 41 AI Act remain a fallback where standards are late. Whether that route is taken for cybersecurity, and when, was open at publication.

The consequences are structured differently as well. The MDR leaves penalties to the Member States under Art. 113 MDR, so an MDR cybersecurity failure is punished, if at all, under national law and through the certificate. The AI Act carries its own scale. Under Art. 99(4)(a) AI Act, a provider that breaches its obligations under Art. 16 AI Act, the first of which is compliance with the requirements that include Art. 15 AI Act, faces administrative fines of up to EUR 15 million or 3 % of total worldwide annual turnover, whichever is higher. Art. 99(4)(b) AI Act applies the same scale to the authorized representative. Art. 44(3) AI Act separately obliges a notified body that finds a system no longer meets the requirements to suspend, withdraw or restrict the certificate, taking account of proportionality, unless corrective action is taken within the deadline the body sets.9AI Act (n 5), Art. 99(4), Art. 16(a), Art. 17(1)(e), Art. 44(3); MDR (n 2), Art. 113. A vulnerability discovered after placement on the market therefore has an AI Act price and an MDR price, and the AI Act's is computed on group turnover.

The timing, and possibly the scope, were rewritten in the weeks before publication. The Digital Omnibus on AI, proposed on 19 November 2025, was adopted by the Council on 29 June 2026 and signed on 8 July 2026. At publication the Official Journal had yet to carry it. As adopted, it moves the application date for Annex I high-risk systems, the route a regulated medical device travels, from 2 August 2027 to 2 August 2028. It also inserts a new Art. 2(13) AI Act, under which the requirements of Art. 9 to Art. 15 and Art. 17 to Art. 25 AI Act may be limited for Art. 6(1) AI Act systems where and to the extent that the Annex I product legislation provides an equivalent or higher level of protection. The specifying delegated acts are due by 2 August 2027.10Digital Omnibus on AI, COM(2025) 836, adopted 29 June 2026: Art. 113, Art. 2(13), Art. 111(2) and Art. 6(1a) to (1c) AI Act as amended. Whether a delegated act will treat sections 17.2 and 17.4 of Annex I MDR as equivalent to Art. 15(5) AI Act is the open question that sits under everything above. The MDR text does not mention the attacks Art. 15(5) AI Act names, and the joint FAQ reads the AI Act's list into the MDR file rather than the other way round. The delegated act had not been drafted at publication, however, and a compliance program built for 2028 may be building to a requirement that a 2027 delegated act narrows, or leaves entirely in place.

Three cybersecurity perimeters around one AI-enabled medical device Three panels around a small central box. The left panel, EU MDR as applied in Switzerland through the MepV, lists Annex I sections 17.2, 17.4 and 23.4(ab), the device in its IT environment, MDCG 2019-16 Rev.1 and its joint-responsibility model, no harmonized cybersecurity standard cited in the Official Journal, penalties left to the Member States, and the exclusion of medical devices from the Cyber Resilience Act. The right panel, EU AI Act on the Annex I route, lists Art. 15(5), Annex IV point 2(h) and Art. 99(4) AI Act, the model and its data pipeline with poisoning and evasion, confidentiality attacks and model flaws, no harmonized standard and a closed Cyber Resilience Act bridge, fines of up to EUR 15 million or 3 % of turnover, and application from 2 August 2028 as adopted. The bottom panel, Switzerland, lists Art. 74b(1)(u) ISG, Art. 24 DSG and Art. 66 MepV, the 24-hour report to the BACS with no Swiss-seat condition, breach notification to the EDÖB and vigilance reporting to Swissmedic, and no AI Act, with a consultation draft due at the end of 2026. The central box, connected to all three panels, reads one device, one notified body, four clocks. Three cybersecurity perimeters, one AI-enabled device EU MDR (in Switzerland: MepV) Annex I, sections 17.2, 17.4, 23.4(ab) The device in its IT environment: networks, access MDCG 2019-16 Rev.1: joint responsibility No harmonized cybersecurity standard cited in the OJ Penalties left to Member States (Art. 113 MDR) Medical devices excluded from the CRA EU AI Act, Annex I route Art. 15(5), Annex IV point 2(h), Art. 99(4) AI Act The model and its data pipeline: poisoning, evasion Confidentiality attacks, model flaws No harmonized standard; CRA bridge closed Fines up to EUR 15 million or 3 % of turnover Applies from 2 Aug 2028 (as adopted) One device one notified body four clocks Switzerland Art. 74b(1)(u) ISG, Art. 24 DSG, Art. 66 MepV 24-hour report to the BACS; no Swiss-seat condition EDÖB breach notification; Swissmedic vigilance report No AI Act; consultation draft due end 2026
Regulatory perimeters reaching the cybersecurity of one AI-enabled medical device on the EU and Swiss markets. Shown are the MDR's device requirements, applied in Switzerland through the MepV; the AI Act's model-level requirements on the Annex I route, the 2 August 2028 date being the adopted but, at publication, unpublished amendment; and the Swiss reporting and data-protection layer.

3. Switzerland's Position: MepV, ISG and the Absent AI Act

Switzerland is not an EU Member State, and neither the MDR nor the AI Act has direct effect there. The Swiss device regime, the MepV, incorporates the EU text rather than restating it. Art. 6(2) MepV requires every device to meet the general safety and performance requirements of Annex I of the EU-MDR, so sections 17.2, 17.4 and 23.4(ab) bind a manufacturer for the Swiss market word for word. Art. 6(4) MepV attaches a presumption of conformity to the technical standards Swissmedic designates.11Medizinprodukteverordnung (MepV) (SR 812.213), Art. 6, Art. 25(4), Art. 51, Art. 66. Swissmedic's information sheet on medical device software, in its April 2026 version, points manufacturers to MDCG 2019-16 for cybersecurity and says nothing about artificial intelligence.12Swissmedic, 'Merkblatt Medizinprodukte-Software' (Version 3.0, 21. April 2026). On the device layer, then, Switzerland is the MDR without the AI Act, and for a US reader the second half of that sentence matters more. A Swiss market entry tests nothing that Art. 15 AI Act will test in the EU.

There is no Swiss AI Act, and none is scheduled. On 12 February 2025 the Federal Council chose sector-specific amendments over a horizontal statute, committed to ratifying the Council of Europe's framework convention on artificial intelligence, which Switzerland signed on 27 March 2025, and ordered a consultation draft by the end of 2026. Nothing in that decision touches the MepV, and healthcare is named only as a sector in which regulatory work continues, a course examined in Insight 61.13Federal Council decision of 12 February 2025; Council of Europe AI convention, signed for Switzerland 27 March 2025. What does reach a Swiss AI-enabled device is the DSG, which the EDÖB declared directly applicable to AI-supported processing in November 2023. Health data are besonders schützenswerte Personendaten under Art. 5 lit. c DSG, and Art. 8 DSG requires data security appropriate to the risk. Art. 24 DSG requires a breach that is likely to result in a high risk for the data subject to be reported to the EDÖB as quickly as possible.14Datenschutzgesetz (DSG) (SR 235.1), Art. 5 lit. c, Art. 8, Art. 24; EDÖB statement of 9 November 2023. A poisoning attack on a model trained on Swiss patient data is, on the Swiss side, a data-security event before it is a device event.

The Swiss layer that has no EU or US counterpart in this form is the ISG. Since 1 April 2025, Art. 74b(1) ISG has listed the authorities and organizations that must report cyberattacks to the BACS, and two of its letters reach an AI-enabled device from both ends. Letter f covers hospitals on the cantonal hospital lists under the KVG. Letter u covers manufacturers of hardware or software whose products are used by critical infrastructures, where the product has a remote-maintenance access or serves the control and monitoring of operational systems and processes or the safeguarding of public safety, with no requirement that the manufacturer have a seat in Switzerland. Art. 74b(3) ISG applies the duty to attacks with effects in Switzerland even where the affected IT systems are abroad. The report is due within 24 hours of discovery under Art. 74e(1) ISG. The triggers in Art. 74d ISG include an attack that led to manipulation or leakage of information or that remained undetected over a longer period, and the size exemptions in Art. 12 CSV contain none for letter u manufacturers. Since 1 October 2025, willful non-compliance with a final BACS order that carries the penal warning is punishable under Art. 74h ISG with a fine of up to CHF 100,000.15Informationssicherheitsgesetz (ISG) (SR 128), Art. 74b, 74d, 74e, 74h; Cybersicherheitsverordnung (CSV) (SR 128.51), Art. 12, Art. 14. The reach of letter u toward a US group with no Swiss entity is examined in Insight 58. The point here is narrower. A connected AI device in a Swiss hospital puts the manufacturer, not only the hospital, on a 24-hour clock that neither the MDR's vigilance system nor the AI Act's serious-incident regime runs.

The clocks then multiply. The same exploited vulnerability may be a schwerwiegendes Vorkommnis occurring in Switzerland, to be reported to Swissmedic under Art. 66 MepV, a duty the Swiss authorized representative carries under Art. 66(2bis) MepV where one is required. It may be a serious incident under Art. 87 MDR for the EU market, and it may call for an Art. 24 DSG notification, which the CSV itself uses as one of the circumstances that make an attack reportable to the BACS. On the AI Act side it may be nothing at all unless the incident infringes obligations under Union law intended to protect fundamental rights, because Art. 73(10) AI Act limits serious-incident reporting for MDR devices to that single category.16AI Act (n 5), Art. 2(1)(a), Art. 22, Art. 73(10); MDR (n 2), Art. 11, Art. 87. Four authorities, four definitions of the event and four moments from which time runs, none coordinated with the others.

Third-country status compounds it. A Swiss manufacturer has been a third-country manufacturer under the MDR since the medical-devices chapter of the Swiss-EU mutual recognition agreement stopped delivering mutual recognition on 26 May 2021. It needs an authorized representative under Art. 11 MDR, who is liable for defective devices on the same basis as, and jointly and severally with, the manufacturer under Art. 11(5) MDR. Under the AI Act the same manufacturer is a provider established in a third country and must, under Art. 22 AI Act, appoint by written mandate an authorized representative established in the Union. That representative's tasks include verifying that the EU declaration of conformity and the technical documentation have been drawn up and that an appropriate conformity assessment procedure has been carried out, and keeping the documentation at the disposal of the authorities for ten years.16 The result is two mandates, with different tasks and different liability, that may or may not sit with the same entity. The protocols to the mutual recognition agreement signed on 2 March 2026 do not, on their own terms, restore the devices chapter, a matter Insight 53 takes up.17Swiss-EU MRA (SR 0.946.526.81), Annex 1, ch 4; protocols signed 2 March 2026; Botschaft of 13 March 2026. In the opposite direction the border is porous. Art. 25(4) MepV treats certificates from EU-designated notified bodies as equivalent to Swiss ones where the procedure applied and the body's qualification can credibly be shown to match the Swiss requirements. In practice the EU body's verdict on the AI Act layer is therefore also the verdict Switzerland receives, from a body Swissmedic did not designate.11

4. Conformity Assessment Under Two Regimes at Once

Art. 43(3) AI Act folds the AI Act's requirements into the MDR conformity assessment. The provider follows the MDR procedure, the requirements of Section 2 of Chapter III AI Act apply and form part of that assessment, and an MDR notified body may control them provided its compliance with the AI Act's notified-body criteria in Art. 31(4), (5), (10) and (11) AI Act was assessed in the notification procedure.18AI Act (n 5), Art. 43(3) and (4), Art. 3(23), Art. 8(2), Art. 25(3) and (4); MDR (n 2), Annex II 6.1(b), Annex IX 4.10. The general shape of that single assessment, and the capacity question behind it, are examined in Insight 51. For cybersecurity the narrower point is what evidence the one assessor will read. An MDR technical file carries software verification and validation under section 6.1(b) of Annex II MDR, a description of the design and development process and evidence of validation in the finished device. An AI Act file carries, under Annex IV AI Act, the cybersecurity measures put in place, the validation and testing data and their characteristics, and the metrics used to measure robustness. Art. 8(2) AI Act lets the provider integrate the second into the first. Integration is a choice about where the evidence sits, not about what evidence exists. A file that documents penetration testing of the device's interfaces without documenting how the training pipeline was protected against poisoning is complete under one regime and silent under the other.

The assessor's competence is the second variable. The joint FAQ expects cybersecurity to be part of the risk management system and the quality management system and therefore subject to conformity assessment.7 The body assessing an adversarial-robustness argument, however, is a medical-device body whose AI Act competence, if assessed at all, was assessed inside the MDR notification procedure. For a Swiss manufacturer the choice of that body is also the choice of the body whose certificate Switzerland accepts under Art. 25(4) MepV. Whether a body designated for the device type, competent in the technology and assessed for the AI Act layer will exist when it is needed is a question the regulation does not answer and the standardization calendar makes harder.

Change control is where the two regimes disagree most visibly. Under Art. 43(4) AI Act a substantial modification requires a new conformity assessment, unless, for a system that continues to learn after being placed on the market, the change was pre-determined at the initial assessment and described in the technical documentation under point 2(f) of Annex IV AI Act. Art. 3(23) AI Act defines the term autonomously, as a change not foreseen or planned in the initial conformity assessment that affects compliance with the Section 2 requirements or modifies the intended purpose. The joint FAQ confirms that the concept is autonomous and not aligned with the MDR's own test, under which section 4.10 of Annex IX MDR requires the notified body's approval for changes that could affect safety and performance or the conditions prescribed for use.18 A model retrained to close a poisoning vulnerability affects compliance with Art. 15 AI Act by construction. Whether it also affects safety and performance under the MDR is a separate question with a separate answer. The MDR side of that dilemma, the security patch that becomes a significant change, is examined in Insight 18. The AI Act adds a second test to the same patch.

The timing rule adds a third. Art. 111(2) AI Act, as amended by the Digital Omnibus on AI, applies the regulation to high-risk systems placed on the market before the Chapter III application date, 2 August 2028 for the device route, only if they are subject to significant changes in their designs from that date. The joint FAQ's Question 31 still reasons from the 2 August 2027 date the AI Act carried when the FAQ was written, so the guidance and the amending act point at different years.19Digital Omnibus on AI (n 10), Art. 111(2) AI Act as amended; AIB 2025-1 / MDCG 2025-6 (n 7), Q29 to Q31. Whether a retrained model, a new feature-extraction layer or a security update is a significant change in design decides whether a device certified in 2027 enters the AI Act at all. The regulation does not define the phrase.

Finally, the model rarely belongs entirely to the manufacturer. Art. 25(3) AI Act makes the product manufacturer the provider where the safety-component AI system is placed on the market together with the product under the manufacturer's name. Art. 25(4) AI Act requires a written agreement with any third party supplying an AI system, tools, services, components or processes, free and open-source tools aside (a carve-out that does not reach general-purpose AI models). The agreement must specify the information, capabilities, technical access and other assistance, based on the generally acknowledged state of the art, that the provider needs for compliance.18 A licensed foundation model, a contract-developed classifier and a pre-trained component bought from a US parent all fall within that provision. The assistance the provider needs for Art. 15(5) AI Act, evidence about training data and pre-trained components that could have been poisoned, is precisely what a licensor is least willing to hand over.

5. Strategic Considerations

The questions that decide exposure here are not the ones a cybersecurity program is built to answer. Whether the device's AI is a safety component or is the device, and whether the notified body's involvement satisfies Art. 6(1) AI Act, determines whether Art. 15 AI Act applies at all. The Digital Omnibus on AI narrows the safety-component definition, excluding systems used solely for non-safety-related assistance, optimization or quality control unless their failure would endanger health and safety, and that narrowing moves the line for devices that were confidently on one side of it.10 Which entity in the group is the provider, the Swiss subsidiary that places the device on the market or the US parent that trained the model, determines who must hold the Art. 25(4) AI Act agreement and whose turnover the fine is computed on. And whether the group has a Swiss seat, or merely a device with a remote-maintenance link into a Swiss hospital, determines whether a 24-hour BACS clock runs beside the Swissmedic, EDÖB and EU vigilance clocks.

Behind those sit questions that depend on facts only the company holds. What the notified body was told at the initial assessment about pre-determined changes decides whether a security retraining is a substantial modification or an anticipated one. That record was written by a regulatory team that may never have been asked about adversarial robustness. What the model license says about training-data provenance decides whether the Art. 15(5) AI Act evidence can be assembled at all. What the incident-response plan treats as discovery decides which of four clocks starts first, and whether a report to one authority shows, by its timing, that a report to another was late.

For a US-listed parent the feedback loop is concrete. A certificate suspended under Art. 44(3) AI Act or an administrative fine computed on group turnover is a candidate for risk-factor disclosure and for a Board's review of its D&O coverage. A BACS order against a Swiss subsidiary is the kind of foreign administrative action that standard US policies sub-limit. None of that appears in a section 524B file.

Whether a given device sits on the right side of each of these lines depends on the intended purpose as written, the notified body's designation, the contracts around the model and the group's Swiss footprint. Those are questions of fact and of drafting, and they require analysis tailored to the device, the entities involved and the commercial context.

REFERENCES

01
Federal Food, Drug, and Cosmetic Act § 524B, 21 U.S.C. § 360n-2 (ensuring cybersecurity of devices), added by section 3305 of the Food and Drug Omnibus Reform Act of 2022, Pub. L. No. 117-328, div. FF, enacted 29 December 2022; § 524B(c) defines a cyber device as a device that includes software validated, installed or authorized by the sponsor as a device or in a device, has the ability to connect to the internet, and contains technological characteristics validated, installed or authorized by the sponsor that could be vulnerable to cybersecurity threats, and § 524B(b)(3) requires a software bill of materials. FDA, 'Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions' (Guidance for Industry and FDA Staff, 3 February 2026, docket FDA-2021-D-1158), superseding the version of 27 June 2025, which had itself updated the original guidance of 27 September 2023. FDA, 'Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations' (Draft Guidance for Industry and FDA Staff, 7 January 2025), whose cybersecurity section lists, as examples of AI risks that cybersecurity threats can impact, data poisoning, model inversion and stealing, model evasion, data leakage, overfitting, model bias and performance drift; a draft, not for implementation.
02
Regulation (EU) 2017/745 of the European Parliament and of the Council of 5 April 2017 on medical devices [2017] OJ L117/1 (MDR), Annex I, sections 17.1 to 17.4 (electronic programmable systems and software that are devices in themselves; section 17.2 on development and manufacture in accordance with the state of the art taking into account the principles of development life cycle, risk management, including information security, verification and validation; section 17.4 on minimum requirements concerning hardware, IT networks characteristics and IT security measures, including protection against unauthorised access, necessary to run the software as intended) and section 23.4(ab) (the same minimum requirements in the instructions for use); Art. 8 (presumption of conformity from harmonised standards whose references have been published in the Official Journal); Art. 113 (rules on penalties laid down by the Member States).
03
MDCG 2019-16 Rev.1, 'Guidance on Cybersecurity for medical devices' (Medical Device Coordination Group, December 2019, revision 1 of July 2020), mapping the MDR's cybersecurity content to Annex I sections 17.2, 17.4 and 23.4(ab) among others, describing a defense-in-depth design and a security risk management process alongside safety risk management, and, in section 2.6, treating manufacturers, suppliers, healthcare providers, patients, integrators, operators and regulators as sharing responsibility for a secured environment, with 'joint responsibility' recorded as the preferred term; the operator is expected to maintain the security of the operating environment and to install the security patches the manufacturer prescribes.
04
Commission Implementing Decision C(2021) 2406 of 14 April 2021 on a standardisation request to the European Committee for Standardization and the European Committee for Electrotechnical Standardization as regards medical devices in support of Regulation (EU) 2017/745 and in vitro diagnostic medical devices in support of Regulation (EU) 2017/746 (M/575), as amended by C(2023) 694 of 31 January 2023 and C(2024) 3371 of 27 May 2024, whose Annex I lists 'Health software and health IT systems safety, effectiveness and security, Part 5-1: Security, Activities in the product life cycle (IEC 81001-5-1)' with a deadline for adoption of 27 May 2028. Commission Implementing Decision (EU) 2021/1182 of 16 July 2021 on the harmonised standards for medical devices drafted in support of Regulation (EU) 2017/745 [2021] OJ L256/100, as amended by Implementing Decisions (EU) 2022/6, 2022/757, 2023/1410, 2024/815, 2024/2631, 2025/681, 2025/2078, 2026/193, 2026/760 and, the last before publication, (EU) 2026/1231 of 11 June 2026. None of those decisions cites EN IEC 81001-5-1:2022, so the standard carried no presumption of conformity under Art. 8 MDR (n 2) as of publication.
05
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (AI Act) [2024] OJ L 2024/1689, Art. 6(1) and Annex I, Section A (the MDR and the IVDR among the listed Union harmonisation legislation); Art. 15(1) (appropriate level of accuracy, robustness and cybersecurity, performed consistently throughout the lifecycle), Art. 15(4) (resilience against errors, faults or inconsistencies; feedback loops in systems that continue to learn) and Art. 15(5) (resilience against attempts by unauthorised third parties to alter use, outputs or performance by exploiting system vulnerabilities; technical solutions appropriate to the relevant circumstances and the risks; measures, where appropriate, against data poisoning, model poisoning, adversarial examples or model evasion, confidentiality attacks or model flaws); Art. 9(2)(a) and (c) (risks to health, safety or fundamental rights; post-market monitoring data); Art. 13(3)(b)(iii) (instructions for use); Art. 17(1)(e) (means to ensure compliance where harmonised standards are not applied in full); Art. 41 (common specifications); Art. 72 (post-market monitoring); Annex IV, point 2(g) and (h) (validation and testing procedures, metrics for accuracy and robustness, and cybersecurity measures put in place).
06
Regulation (EU) 2024/2847 of the European Parliament and of the Council of 23 October 2024 on horizontal cybersecurity requirements for products with digital elements (Cyber Resilience Act) [2024] OJ L 2024/2847, Art. 2(2)(a) (the Regulation does not apply to products with digital elements to which Regulation (EU) 2017/745 applies), Art. 12(1) (products within the CRA's scope that are classified as high-risk AI systems under Art. 6 AI Act and fulfil the essential cybersecurity requirements of Part I of Annex I CRA are deemed to comply with the cybersecurity requirements of Art. 15 AI Act, subject to the conditions in points (a) to (c)); Art. 71 (application from 11 December 2027, with the reporting obligations of Art. 14 from 11 September 2026).
07
Medical Device Coordination Group and European Artificial Intelligence Board, 'Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA)' (AIB 2025-1 / MDCG 2025-6, 19 June 2025), Q22 (cybersecurity measures required by the AIA and the MDR/IVDR: technical solutions to address AI-specific vulnerabilities; securing AI-specific assets such as training data sets or the trained model; preventing data and model poisoning; cybersecurity as part of the risk management system and the quality management system and therefore subject to conformity assessment; and the note that medical devices and in vitro diagnostic medical devices are outside the scope of the Cyber Resilience Act), Q29 (substantial modification as an autonomous concept under the AIA) and Q30 (pre-determined changes). The document states that the views expressed in it are not legally binding and that only the Court of Justice of the European Union can give binding interpretations of Union law.
08
Commission Implementing Decision C(2025) 3871 final of 23 June 2025 on a standardisation request to the European Committee for Standardisation and the European Committee for Electrotechnical Standardisation as regards high-risk AI systems in support of Regulation (EU) 2024/1689 and repealing Implementing Decision C(2023) 3215, Art. 1 (harmonised standards and European standardisation deliverables to be drafted by 31 August 2025), Annex I (the requested deliverables, including harmonised standards on cybersecurity specifications for AI systems) and Art. 5 (expiry on 28 February 2027). European Commission, 'Understanding the standardisation of the AI Act' (Shaping Europe's digital future, version of 10 March 2026), stating that the first harmonised standards were expected to be published by CEN and CENELEC in 2026 and that the Commission would then assess them for citation in the Official Journal. No harmonized standard supporting Art. 15 AI Act (n 5) had been cited in the Official Journal as of publication.
09
AI Act (n 5), Art. 99(4)(a) and (b) (administrative fines of up to EUR 15 000 000 or, for an undertaking, up to 3 % of its total worldwide annual turnover for the preceding financial year, whichever is higher, for non-compliance with the obligations of providers under Art. 16 and of authorised representatives under Art. 22), Art. 16(a) (providers shall ensure that their high-risk AI systems are compliant with the requirements set out in Section 2), Art. 17(1)(e) and Art. 44(3) (suspension, withdrawal or restriction of a certificate where the system no longer meets the requirements, unless compliance is ensured by corrective action within the deadline set by the notified body); MDR (n 2), Art. 113.
10
European Commission, Proposal for a Regulation of the European Parliament and of the Council amending Regulations (EU) 2024/1689 and (EU) 2018/1139 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI), COM(2025) 836 final (19 November 2025), procedure file 2025/0359(COD). A provisional interinstitutional agreement was reached on 7 May 2026; the European Parliament adopted its first-reading position on 16 June 2026; the Council adopted the act on 29 June 2026; the final act was signed on 8 July 2026 and, as of publication, awaited publication in the Official Journal, entering into force on the third day following it. As adopted, the amending regulation replaces Art. 113, third paragraph, point (c) AI Act (n 5) so that Chapter III, Sections 1 to 3 apply from 2 December 2027 to Annex III high-risk systems and from 2 August 2028 to systems classified as high-risk under Art. 6(1) and Annex I; inserts Art. 2(13) AI Act, under which the application of specific requirements or obligations in Art. 9 to Art. 15 and Art. 17 to Art. 25 may be limited for Art. 6(1) systems where and to the extent that the Annex I, Section A legislation provides an equivalent or higher level of protection of health, safety or fundamental rights and the limitation does not reduce the overall level of protection, the Commission to adopt the specifying delegated acts by 2 August 2027; amends Art. 111(2) AI Act so that the regulation applies to high-risk systems placed on the market or put into service before the date of application of Chapter III only if they are subject to significant changes in their designs from that date; and inserts Art. 6(1a) to (1c) AI Act on systems that do not qualify as safety components.
11
Medizinprodukteverordnung (MepV) vom 1. Juli 2020 (SR 812.213), Art. 6(2) (devices must satisfy the general safety and performance requirements of Anhang I EU-MDR), Art. 6(4) (presumption of conformity for devices in accordance with the technical standards or common specifications designated by Swissmedic), Art. 25(4) (certificates issued by bodies designated under EU law with a seat in an EU or EEA state, and not recognised under an international agreement, are treated as equivalent to certificates of Swiss bodies where it can credibly be shown that the conformity assessment procedures applied satisfy the Swiss requirements and that the issuing body has a qualification equivalent to that required in Switzerland; inserted with effect from 26 May 2021), Art. 51 (Swiss authorised representative), Art. 66(1) (reporting of serious incidents occurring in Switzerland to Swissmedic) and Art. 66(2bis) (where a representative is required under Art. 51, that representative bears responsibility for the report, the transfer of the duty to be agreed in writing in the mandate).
12
Swissmedic, 'Merkblatt Medizinprodukte-Software' (Merkblatt BW630_30_007, Version 3.0, 21. April 2026), restating that devices must meet the general safety and performance requirements of Anhang I EU-MDR and that the requirements for programmable electronic systems include IT security and development in accordance with the state of the art, taking into account the software life cycle, risk management including information security, verification and validation, and referring manufacturers to MDCG 2019-16 (n 3) as guidance on meeting the cybersecurity-related requirements of Anhang I.
13
Schweizerischer Bundesrat, 'KI-Regulierung: Bundesrat will Konvention des Europarats ratifizieren' (Medienmitteilung, 12. Februar 2025), deciding that Switzerland should ratify the Council of Europe's convention on artificial intelligence and make the necessary adjustments to Swiss law, that legislative adjustments should be sector-specific where possible, and that the EJPD, with the UVEK and the EDA, should prepare a consultation draft by the end of 2026; the release names healthcare among the areas in which sectoral regulatory work continues. Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225, adopted 17 May 2024, opened for signature at Vilnius on 5 September 2024), signed for Switzerland on 27 March 2025 and, as of publication, not ratified by Switzerland.
14
Bundesgesetz über den Datenschutz (Datenschutzgesetz, DSG) vom 25. September 2020 (SR 235.1), in force 1 September 2023, Art. 5 lit. c (health data among the besonders schützenswerte Personendaten), Art. 8 (data security appropriate to the risk, through suitable technical and organizational measures) and Art. 24 (notification to the EDÖB, as quickly as possible, of a breach of data security likely to result in a high risk for the personality or fundamental rights of the data subject; under Art. 24(5bis), inserted with effect from 1 April 2025 by the same federal act that introduced the ISG reporting duty (n 15), the EDÖB may forward the notification to the Bundesamt für Cybersicherheit with the controller's consent). EDÖB, 'Geltendes Datenschutzgesetz ist auf KI direkt anwendbar' (9. November 2023), stating that the DSG is formulated in a technology-neutral manner and is therefore directly applicable to AI-supported data processing.
15
Bundesgesetz über die Informationssicherheit (Informationssicherheitsgesetz, ISG) vom 18. Dezember 2020 (SR 128), Art. 74a to 74h, inserted by the federal act of 29 September 2023 introducing a duty to report cyberattacks on critical infrastructures, in force since 1 April 2025 (Art. 74g and 74h since 1 October 2025): Art. 74b(1)(f) (hospitals on the cantonal hospital list under Art. 39(1)(e) KVG), Art. 74b(1)(u) (manufacturers of hardware or software whose products are used by critical infrastructures, where the hardware or software has a remote-maintenance access or serves the control and monitoring of operational systems and processes or the safeguarding of public safety; no Swiss-seat condition), Art. 74b(3) (application to cyberattacks with effects in Switzerland even where the affected IT systems are located abroad), Art. 74d (reportable attacks: those that endanger the functioning of the critical infrastructure, have led to manipulation or leakage of information, remained undetected over a longer period, or are connected with extortion, threats or coercion), Art. 74e(1) (report within 24 hours of discovery), Art. 74g (notice and deadline, then a formal order with a new deadline and reference to the penalty) and Art. 74h(1) (fine of up to CHF 100,000 for willfully failing to comply with a final BACS order issued with reference to that penalty). Cybersicherheitsverordnung (CSV) vom 7. März 2025 (SR 128.51), in force 1 April 2025, Art. 12 (exemptions by size or threshold for organizations under letters a, d, g, h, l, m, n, p and t of Art. 74b(1) ISG, none for letter u) and Art. 14(2)(b) (an attack counts as having led to manipulation or leakage of information where, among other circumstances, a notification under Art. 24 DSG has been made).
16
AI Act (n 5), Art. 2(1)(a) (application to providers placing AI systems on the market in the Union irrespective of whether they are established in the Union or in a third country), Art. 22(1) (providers established in third countries shall, prior to making their high-risk AI systems available on the Union market, appoint by written mandate an authorised representative established in the Union) and Art. 22(3)(a) and (b) (verification that the EU declaration of conformity and the technical documentation have been drawn up and that an appropriate conformity assessment procedure has been carried out; keeping the technical documentation, the declaration and any certificate at the disposal of the competent authorities for ten years after placing on the market or putting into service); Art. 73(10) (for high-risk AI systems that are devices, or safety components of devices, covered by the MDR or the IVDR, the notification of serious incidents is limited to those referred to in Art. 3, point (49)(c), the infringement of obligations under Union law intended to protect fundamental rights). MDR (n 2), Art. 11(1) (sole authorised representative where the manufacturer is not established in a Member State), Art. 11(5) (the authorised representative is legally liable for defective devices on the same basis as, and jointly and severally with, the manufacturer) and Art. 87(1) (reporting of serious incidents and field safety corrective actions).
17
Agreement between the European Community and the Swiss Confederation on mutual recognition in relation to conformity assessment (signed 21 June 1999, entered into force 1 June 2002) [2002] OJ L114/369 (SR 0.946.526.81), Annex 1, ch 4 (medical devices), whose trade-facilitating effects for devices under the MDR (n 2) ceased on 26 May 2021. Schweizerischer Bundesrat, Botschaft über das Paket «Stabilisierung und Weiterentwicklung der Beziehungen Schweiz-EU (Bilaterale III)» (13. März 2026), Geschäft 26.023, BBl 2026 615, recording the signature of the bulk of the package in Brussels on 2 March 2026, including the amending protocol (BBl 2026 619) and the institutional protocol (BBl 2026 620) to the agreement; the realignment of the medical-devices chapter with the MDR is left to a later decision of the Committee established under the agreement, and no published instrument dated the protocols' entry into force as of publication.
18
AI Act (n 5), Art. 43(3) (for high-risk AI systems covered by the Annex I, Section A legislation, the provider follows the conformity assessment procedure required under those acts; the Section 2 requirements are part of that assessment; notified bodies notified under those acts may control conformity with the Section 2 requirements provided that their compliance with Art. 31(4), (5), (10) and (11) has been assessed in the context of the notification procedure under those acts), Art. 43(4) (new conformity assessment in the event of a substantial modification; pre-determined changes to a continuously learning system described under point 2(f) of Annex IV are not substantial modifications), Art. 3(23) (definition of substantial modification), Art. 8(2) (the provider's choice of integrating the AI Act testing, reporting, information and documentation into the documentation and procedures already required under the Annex I legislation), Art. 25(3) (the product manufacturer as provider of a high-risk AI system that is a safety component placed on the market or put into service under its name or trademark) and Art. 25(4) (written agreement between the provider and a third party supplying an AI system, tools, services, components or processes, specifying the necessary information, capabilities, technical access and other assistance based on the generally acknowledged state of the art; the exception for tools, services, processes or components made accessible to the public under a free and open-source licence does not extend to general-purpose AI models). MDR (n 2), Annex II, section 6.1(b) (software verification and validation, describing the software design and development process and evidence of the validation of the software as used in the finished device) and Annex IX, section 4.10 (changes to the approved device require approval from the notified body where they could affect the safety and performance of the device or the conditions prescribed for its use).
19
Digital Omnibus on AI (n 10), Art. 111(2) AI Act (n 5) as amended. The amended paragraph retains the public-authority backstop: in any case, providers and deployers of high-risk AI systems intended to be used by public authorities must take the necessary steps to comply by 2 August 2030. AIB 2025-1 / MDCG 2025-6 (n 7), Q29 (substantial modification is an autonomous concept under the AIA, defined in Art. 3(23), and is not aligned with the changes that may require a new conformity assessment under the MDR/IVDR), Q30 (pre-determined changes described under point 2(f) of Annex IV are not substantial modifications) and Q31 (a high-risk medical-device AI system already on the market that undergoes a significant change in its design on or after 2 August 2027 becomes subject to the AI Act obligations), the last of which reasons from the application date the AI Act carried before its amendment.

Where one AI-enabled device answers to the MDR, the AI Act and Swiss law at once, the allocation of its cybersecurity duties across entities, files and reporting clocks is a matter for tailored analysis.

Get in Touch