Engineer working on a laptop connected to a device test rig, surrounded by wiring harnesses and test hardware

Firmware vulnerabilities & fixes

ISO/IEC 42001:2023: the AI management system (AIMS) standard, explained

Last updated:

ISO/IEC 42001:2023 is the world's first certifiable international standard for an artificial intelligence management system (AIMS), built around 38 Annex A controls grouped under nine control objectives on top of the same Harmonized Structure clauses (4 through 10) used by ISO/IEC 27001. Published 18 December 2023 by ISO and IEC, it sets auditable requirements for how an organization governs the AI systems it develops, provides or uses, covering risk, lifecycle control and impact on people and society. It is best understood as "27001's sibling for AI governance" rather than a technical AI-safety specification.

Clause titles, annex status and scope below are taken from the publisher's free ISO/IEC 42001:2023 preview (cover, contents, introduction and clauses 1–3). The Annex A control names come from a crosswalk listed on NIST's AI Resource Center and drafted against the final draft (FDIS) text, so check them against the purchased standard before a compliance decision.

Key takeaways

  • Clause titles, annex status and scope below are taken from the publisher's free ISO/IEC 42001:2023 preview (cover, contents, introduction and clauses 1–3).
  • The official contents list shows where the AI-specific sub-clauses sit inside the shared structure.
  • Annex A is the standard's normative control catalogue.
  • This is the clause that most clearly separates ISO/IEC 42001 from a conventional information-security management system, and it is the piece an ISO/IEC 27001-certified organization cannot simply reuse.

What does ISO/IEC 42001 actually require?

ISO/IEC 42001 uses ISO's Harmonized Structure — the same template behind ISO/IEC 27001:2022 and ISO 9001 — with normative requirements running through Clauses 4 to 10 as a Plan-Do-Check-Act cycle: Context of the Organization (4), Leadership (5), Planning (6), Support (7), Operation (8), Performance Evaluation (9) and Improvement (10). Clause 6 is where the AI-specific work concentrates: an AI risk assessment and risk treatment process, AI objectives, and the establishment of an AI system impact assessment process. Clause 8 then requires operating and controlling the AI lifecycle against that planning, and Clause 9 requires internal audits and management review producing evidence an external auditor can sample. Because the skeleton is shared with ISO/IEC 27001, an organization already running an ISMS typically reuses its document-control, internal-audit and management-review machinery rather than building a second parallel system from scratch — the new work is almost entirely in the AI-specific content, not the management-system scaffolding. The standard's only normative reference is ISO/IEC 22989:2022, the AI concepts and terminology standard.

Which clauses carry the AI-specific requirements?

The official contents list shows where the AI-specific sub-clauses sit inside the shared structure. The introduction says the standard "applies the harmonized structure (identical clause numbers, clause titles, text and common terms and core definitions)", so everything not named in the right-hand column reads much like ISO/IEC 27001.

Clause Shared management-system content AI-specific sub-clauses
4 Context Context, interested parties, management system 4.3 scope of the AI management system
5 Leadership Commitment, roles 5.2 AI policy
6 Planning Risks and opportunities, objectives, changes 6.1.2 AI risk assessment; 6.1.3 AI risk treatment; 6.1.4 AI system impact assessment; 6.2 AI objectives
7 Support Resources, competence, awareness, communication, documented information —
8 Operation Operational planning and control 8.2 AI risk assessment; 8.3 AI risk treatment; 8.4 AI system impact assessment (carried out, not only defined)
9 Performance evaluation Monitoring, internal audit (9.2), management review (9.3) —
10 Improvement Continual improvement, nonconformity and corrective action —

Two annexes are normative: Annex A (reference control objectives and controls) and Annex B (implementation guidance for AI controls). Annex C (organizational objectives and risk sources) and Annex D (use across domains or sectors) are informative.

What's in Annex A, and is a device maker expected to implement all of it?

Annex A is the standard's normative control catalogue. Corroborated secondary reporting — not the paywalled standard text itself — describes 38 controls organized under nine control objectives, A.2 through A.10: AI policies, internal organization, resources for AI systems, assessing impacts of AI systems, the AI system life cycle, data for AI systems, information for interested parties of AI systems, use of AI systems, and third-party and customer relationships. As with ISO/IEC 27001's own Annex A, the standard does not require implementing all 38 controls by default — an organization runs its Clause 6 risk assessment and AI system impact assessment first, then selects applicable controls into a Statement of Applicability and documents a justification for any it excludes. Annex B gives implementation guidance control-by-control, Annex C catalogues AI-specific risk sources (bias, explainability gaps, model drift, safety) that feed the risk assessment, and Annex D addresses sector-specific applicability.

What are the 38 Annex A controls?

The NIST AI RMF to ISO/IEC 42001 crosswalk, written by Microsoft and listed on NIST's crosswalk page (which states that inclusion "does not imply NIST endorsement"), cites 38 distinct controls by their Annex B guidance numbers. They line up with the nine objectives:

Objective Controls Control titles (crosswalk wording)
A.2 AI policies 3 AI policy; alignment with other organizational policies; review of the AI policy
A.3 Internal organization 2 AI roles and responsibilities; reporting of concerns
A.4 Resources for AI systems 5 Resource documentation; data, tooling, system and computing, and human resources
A.5 Assessing impacts of AI systems 4 Impact assessment process; documentation of assessments; impact on individuals and groups; societal impacts
A.6 AI system life cycle 9 Objectives and processes for responsible development; requirements and specification; design and development documentation; verification and validation; deployment; operation and monitoring; technical documentation; event logs
A.7 Data for AI systems 5 Data for development; acquisition; quality; provenance; preparation
A.8 Information for interested parties 4 System documentation and information for users; external reporting; communication of incidents; information for interested parties
A.9 Use of AI systems 3 Processes for responsible use; objectives for responsible use; intended use
A.10 Third-party and customer relationships 3 Allocating responsibilities; suppliers; customers

For a firmware team, objectives A.6 and A.7 are the ones closest to existing engineering work: verification and validation, event logs, technical documentation and data provenance are things a device programme already produces in some form.

Why does the AI system impact assessment matter more than it first looks?

This is the clause that most clearly separates ISO/IEC 42001 from a conventional information-security management system, and it is the piece an ISO/IEC 27001-certified organization cannot simply reuse. Centered on Annex A.5 and required by Clause 8.4, it is not a risk assessment in the security sense — it evaluates organizational impact, impact on individuals or groups, and societal impact, including whether an AI system's selection or scoring criteria could produce biased outcomes against a particular community, and it has to be performed and documented at scheduled intervals and whenever a system changes materially. For a vendor that embeds an on-device model — object detection in a camera, anomaly scoring in an industrial sensor, voice processing in an appliance — this is the clause asking a different question than a firmware security assessment does: not just "can this model be attacked or does it leak data," covered by work like an AI-BOM for embedded AI devices or threat-modeled against MITRE ATLAS, but "who could this model's decisions affect, and was that assessed before and after shipping."

How does certification actually work, and does it satisfy the EU AI Act?

Certification is voluntary and performed by accredited certification bodies under ISO/IEC 17021-1, supplemented by ISO/IEC 42006:2025's AI-specific requirements for the certifiers themselves; national accreditation bodies such as ANAB (US) and UKAS (UK) accredit those certifiers, with ANAB reported to have accredited multiple US certification bodies and UKAS confirming its first accredited body by November 2025. The process follows the same two-stage pattern as other management-system certifications — a Stage 1 documentation review, then a Stage 2 operational audit sampling real evidence that the AIMS runs as documented — leading to a three-year certificate with annual surveillance audits and a full recertification audit at the end of the cycle.

Certification does not, by itself, satisfy the EU AI Act. Multiple legal analyses of the Act's Article 40 mechanism agree that ISO/IEC 42001 has not been cited in the EU Official Journal and therefore does not create a rebuttable legal presumption of conformity; that presumption is reserved for harmonized standards CEN-CENELEC's JTC 21 publishes, with a dedicated standard (EN 18286) specifically targeting the Article 17 quality-management-system duties the Act imposes on high-risk AI providers. Where the two overlap, ISO/IEC 42001 certification is widely treated as strong due-diligence evidence of a working governance program — not a substitute for the Act's own conformity-assessment route. Against the non-certifiable NIST AI Risk Management Framework, the standards are complementary rather than competing: a community-submitted crosswalk on NIST's own Trustworthy and Responsible AI Resource Center maps AI RMF's GOVERN/MAP/MEASURE/MANAGE subcategories to specific 42001 clauses and Annex B controls, and ISO/IEC 42001's own risk-assessment clause draws on the separate guidance-only ISO/IEC 23894:2023 (published 6 February 2023), which is explicitly not designed for standalone certification.

One 2026 change is easy to misread. CEN approved EN ISO/IEC 42001:2026 on 13 March 2026, taking the ISO text over unchanged through CEN-CENELEC/JTC 21, with national adoption due by September 2026. Becoming a European Standard is not the same as being a harmonized standard under the AI Act: the European foreword in the publisher's preview does not mention Regulation (EU) 2024/1689 and has no Annex ZA, the annex harmonized standards use to link clauses to legal requirements. The answer above does not change.

What does ISO/IEC 42006 require of the certification body?

ISO/IEC 42006 matters to buyers because it decides how much a certificate is worth. Its scope, in the final draft preview, "specifies additional requirements to ISO/IEC 17021-1" for bodies certifying an AIMS. Three FDIS requirements stand out:

  • Impartiality (5.2.2). Certification bodies "shall not provide consulting for management systems related to artificial intelligence, information security, data protection … or risk management" to their 42001 clients, and shall not run the client's internal audits under another name.
  • Liability (5.3.2). The body needs insurance or an equivalent, with cover in proportion to the turnover of the clients it audits.
  • Competence (7.1). Audit teams need defined knowledge of AI, AI-related legal obligations and the client's "technical area", which the FDIS describes by AI field (for example computer vision or NLP) and application (for example automated vehicles or healthcare).

When you choose a certifier, ask which technical areas its accreditation covers and whether its auditors have audited an embedded or edge AI scope before.

Worked example: what could a camera maker's statement of applicability look like?

Synthetic example, not a real company or certificate. A maker of smart security cameras trains a person-detection model in-house, ships it in firmware, and licenses a third-party face-matching model for one product line.

  1. Scope (Clause 4.3)Model development and the firmware release process are in scope; the customers' own use of video is outside the company's control and is recorded as a customer-side responsibility.
  2. Impact assessment (6.1.4, 8.4)The team assesses false-positive alarms on individuals, and uneven detection across skin tones and lighting conditions, and re-runs it whenever the model is retrained.

Statement of Applicability (Clause 6.1.3), excerpt:

Control area Decision Evidence the team already has or must create
A.5 impact assessment Applied Assessment record per model version
A.6 verification and validation Applied Test-set results per release, including low-light set
A.6 event logs Applied On-device detection log format and retention
A.7 data provenance Applied Dataset register with source and licence per dataset
A.10 suppliers Applied Contract terms and model card for the licensed face-matching model
A.9 intended use Applied Product documentation stating unsupported uses

Link to security work. The same release record carries the firmware SBOM and the AI-BOM, so an auditor sampling one release sees the model, its data and the software around it together.

What are the common pitfalls?

  • Scoping only the cloud. If a model ships in firmware, the release process that builds and signs it belongs in the AIMS scope.
  • Copying the 27001 risk register. The impact assessment asks who is affected by outputs, not only what can be attacked.
  • Excluding supplier controls because "we did not train it". A licensed or open-weight model is exactly what the third-party objective covers.
  • One impact assessment for every model version. Clause 8.4 expects reassessment when the system changes.
  • Treating certification as AI Act conformity. See the EU AI Act section above.

Where does this fit for a device or firmware vendor?

A device or firmware maker that builds or ships its own AI models — not just one that consumes a third-party model API — sits at the intersection of several of these frameworks at once: a Cyber Resilience Act security baseline for the firmware itself, an AI-BOM for the embedded model's provenance, a threat model against adversarial AI attack techniques, and now a management-system layer asking whether the organization actually governs that model's risk and impact on an ongoing basis rather than once at ship time. ISO/IEC 42001 is the standard that formalizes the last piece — it does not replace firmware or product security work, it sits above it as the accountability and audit-trail layer. For generative features, the NIST AI 600-1 profile adds risk categories you can feed into the Clause 6 assessment.

SoftWSP does not issue ISO/IEC 42001 certificates, audit an AIMS or assess models. What it does cover is the software around the model: the platform extracts a firmware SBOM from the binary as CycloneDX 1.6 JSON, matches the components against NVD, CISA KEV, FIRST EPSS and CISA ICS advisories, and drafts a cited report that an engineer reviews. That report can sit in the same release record as your AI-BOM and test results; see device and firmware security.

This guide cross-checks the publisher's preview of ISO/IEC 42001:2023, the EN ISO/IEC 42001:2026 and ISO/IEC FDIS 42006 previews, and the NIST-listed crosswalk against ISO's own standard pages and the ANSI/ANAB accreditation-body blog. The full Annex A and Annex B text is in the purchased standard, which is the authoritative source for a certification decision. This guide is not a substitute for the standard itself, for ISO/IEC 42006's certification-body requirements, or for legal advice on EU AI Act conformity.

Put this guide to work

Frequently asked questions

What is ISO/IEC 42001?

ISO/IEC 42001:2023, titled "Information technology — Artificial intelligence — Management system," is the world's first certifiable international standard for an AI management system (AIMS). Published 18 December 2023 (Edition 1, dated 2023-12) by ISO and IEC jointly, it was developed by ISO/IEC JTC 1/SC 42, the joint technical committee responsible for AI standardization, after work that began around August 2021. Like ISO/IEC 27001 for information security, it specifies auditable requirements for establishing, implementing, maintaining and continually improving a management system — in this case one governing how an organization develops, provides or uses AI systems, covering risk, impact on people and society, and control across the AI lifecycle, rather than prescribing specific AI techniques or model architectures.

How is ISO/IEC 42001 structured, and does it reuse ISO/IEC 27001's format?

Yes — ISO/IEC 42001 follows ISO's Harmonized Structure (the same template used by ISO/IEC 27001:2022 and ISO 9001), with normative requirements in Clauses 4 through 10: Context of the Organization (4), Leadership (5), Planning (6), Support (7), Operation (8), Performance Evaluation (9) and Improvement (10), run as a Plan-Do-Check-Act cycle. Clauses 1-3 (scope, normative references, terms) establish nothing certifiable on their own. Because it shares a skeleton with ISO/IEC 27001:2022 and ISO 9001, organizations already certified to one of those standards typically find clauses like leadership commitment, documented information and internal audit familiar; the AI-specific content sits mainly in Clause 6 (AI risk assessment and AI system impact assessment), Clause 8 (operational control of the AI lifecycle) and Annex A. It is not derived from ISO/IEC 27001 — the two are sibling standards built on a common template, which is also why ISO sells them together in an "AI and information security management" bundle.

What do Annex A, B and C actually require?

Annex A is the normative control catalogue: per widely corroborated secondary reporting (treat the control count as cross-checked against secondary sources rather than the purchased PDF), it lists 38 controls grouped under nine control objectives, A.2 through A.10 — covering AI policies, internal organization, resources for AI systems, assessing impacts of AI systems, the AI system life cycle, data for AI systems, information for interested parties, use of AI systems, and third-party/customer relationships. As with ISO/IEC 27001's Annex A, organizations select applicable controls via a Statement of Applicability driven by their own risk and impact assessments, documenting a justification for any control they exclude, rather than implementing all 38 by default. Annex B supplies implementation guidance for each Annex A control, and Annex C catalogues AI-specific organizational objectives and risk sources — such as bias, lack of explainability, model drift and safety — that feed Clause 6's risk assessment. A further Annex D addresses applicability across specific sectors.

What is an AI system impact assessment, and why is it the newest part for ISO/IEC 27001-certified organizations?

AI system impact assessment (centered on Annex A.5 and required under Clause 8.4) is one of the most distinctive — and, per multiple implementer write-ups, most frequently misjudged — parts of ISO/IEC 42001, and it has no direct ISO/IEC 27001 equivalent. Unlike a security risk assessment, it evaluates three separate dimensions: organizational impact, impact on individuals or groups, and broader societal impact, including whether an AI system's criteria could produce biased or discriminatory outcomes against a particular community. The standard requires organizations to actually perform and document this assessment at defined intervals and whenever an AI system changes materially, retaining the results as evidence an auditor can sample — a companion standard, ISO/IEC 42005, provides more detailed guidance specifically on how to conduct it. For a device or firmware vendor, this is the clause that extends governance beyond "is the model secure" into "who could this model's decisions or outputs harm, and how was that assessed before and after deployment."

Does ISO/IEC 42001 certification satisfy the EU AI Act, and how does it relate to NIST's AI RMF and ISO/IEC 23894?

No — treat this as a governance foundation, not a legal shortcut. Multiple analyses of the EU AI Act's Article 40 mechanism agree that ISO/IEC 42001 has not been cited in the Official Journal of the European Union and does not by itself create a legal presumption of conformity; that presumption is reserved for harmonized standards CEN-CENELEC's JTC 21 publishes under Article 40, with a dedicated deliverable (EN 18286) specifically targeting the Article 17 quality-management-system duties for high-risk AI providers. Where ISO/IEC 42001's scope does overlap the Act's requirements, certification is widely described as useful due-diligence evidence, not proof of compliance. Against the voluntary, non-certifiable NIST AI RMF 1.0, ISO/IEC 42001 is the complementary auditable counterpart: a community-submitted NIST crosswalk (hosted on NIST's own Trustworthy and Responsible AI Resource Center, with NIST explicitly disclaiming endorsement) maps AI RMF's GOVERN/MAP/MEASURE/MANAGE subcategories to 42001 clauses and Annex B controls. ISO/IEC 42001 also sits alongside ISO/IEC 23894:2023 (published 6 February 2023), a guidance-only AI risk management standard built on ISO 31000's structure that is explicitly not designed for standalone certification — ISO/IEC 42001's Clause 6 risk assessment is where 23894's guidance gets put into a certifiable system.

Sources

  1. ISO — ISO/IEC 42001:2023 standard page (title, edition, scope)
  2. ISO — ISO/IEC 42001 explained (ISO Insights)
  3. ISO — ISO/IEC 23894:2023 standard page (AI risk management guidance)
  4. ANSI/ANAB Blog: ISO/IEC 42001 — Artificial Intelligence Management Systems (AIMS)
  5. ANSI Blog: ISO/IEC 23894:2023 — Guidance on Risk Management for AI
  6. ISO/IEC 42001:2023 — publisher's preview (contents, introduction, clauses 1-3), via iTeh Standards
  7. EN ISO/IEC 42001:2026 — catalogue page (CEN approval 13 March 2026, CEN-CENELEC/JTC 21)
  8. ISO/IEC FDIS 42006 — preview (scope, impartiality, liability, competence)
  9. NIST AI Resource Center — AI RMF crosswalk documents (listing and non-endorsement statement)
  10. NIST AI RMF to ISO/IEC 42001 crosswalk (Microsoft, FDIS-based), hosted on NIST AIRC
Was this guide helpful?

In this cluster

ISO/IEC 42001:2023: the AI management system (AIMS) standard, explained