Skip to main content

Standards & Frameworks — Detail and Relationships

Where 02. Guidelines is "a list and summary," this document is for studying each standard in depth and how they fit together.

:::tip First — "there is no single main standard" AI security/governance has no one standard that rules everything. Different roles stack in layers. "Which is the main one" depends on which layer you mean. :::

I. The map — how the standards relate​

Layer map​

LayerRoleRepresentativeNature
Law/regulationWhat you must obeyEU AI Act · (Korea) AI Framework Act · PIPAMandatory
CertificationThird-party attestationISO/IEC 42001Voluntary (certifiable)
Governance frameworkThe backbone/common language of risk managementNIST AI RMFVoluntary
Risk methodologyHow to identify/assess riskISO/IEC 23894Guidance (not certifiable)
Secure developmentLifecycle security principlesGoogle SAIF · CISA/NCSC guideGuidance
Threat checklistsActual assessment itemsOWASP LLM Top 10 · MITRE ATLASGuidance (hands-on)

The order you weave them in practice​

[frame it]       Build the governance/risk backbone with NIST AI RMF
   │
   ├─[fill method]   Reinforce risk identification/assessment with ISO/IEC 23894
   │
   ├─[assess]        Assess real vulnerabilities with OWASP LLM Top 10 · MITRE ATLAS
   │
   ├─[certify]       Obtain ISO/IEC 42001 certification if required
   │
   └─[comply]        Map to EU AI Act · AI Framework Act · PIPA (mandatory)

Crosswalks (mutual mapping)​

The standards translate into one another. NIST officially provides an AI RMF ↔ ISO/IEC 23894 crosswalk, and when preparing ISO 42001 you can map material organized in NIST RMF language directly. In other words, learn one well and it's easy to carry over to the others.

:::info One-line summary

  • Most used when actually assessing → OWASP LLM Top 10
  • Most referenced as a governance framework → NIST AI RMF
  • What you must obey → law (in Korea, the AI Framework Act and PIPA)
  • Proof by certification → ISO/IEC 42001 :::

II. Governance / risk frameworks​

NIST AI RMF (AI 100-1)​

  • By/year: US NIST, January 2023. Free, voluntary, general-purpose.
  • Structure — four functions: GOVERN (foundation: culture, policy, roles) wraps the rest, and within it MAP (context/risk identification) → MEASURE (measurement/testing) → MANAGE (response) iterate.
  • 7 characteristics of trustworthy AI: valid & reliable / safe / secure & resilient / accountable & transparent / explainable & interpretable / privacy-enhanced / fair (bias managed).
  • Generative-AI companion — GenAI Profile (AI 600-1, Jul 2024): maps ~12 GenAI-specific risks (hallucination, information integrity, data privacy, cybersecurity, CBRN info, harmful bias, IP, etc.) to ~200 suggested actions across the four functions. If you work with LLMs, this Profile is more practical than the core.
  • Nature: no legal force. Gives "what," but the organization fills in "how."
  • Original: NIST AI RMF (official) · GenAI Profile publication — free PDF download

ISO/IEC 23894 — AI risk management (guidance)​

  • By/year: ISO/IEC, 2023. A non-certifiable guidance document.
  • Content: extends the general risk-management standard ISO 31000 to the AI context. Methods to identify/assess/treat AI-specific risks like bias, model drift, explainability, and societal impact.
  • In practice: the methodology that fills the "risk process" ISO 42001 requires. But organizations rarely "implement 23894" directly; NIST AI RMF more often takes that role. Treat it as secondary/reference.
  • Original: ISO/IEC 23894:2023 (catalog) — paid, copyrighted (cannot be attached)

III. Management system / certification​

ISO/IEC 42001 — AI Management System (AIMS)​

  • By/year: ISO/IEC, December 2023. The only certifiable AI standard.
  • Structure: the same PDCA (Plan-Do-Check-Act) management-system frame as ISO 27001 + Annex A controls (AI policy, resources, impact assessment, data/model lifecycle, etc.).
  • Position: does for AI what ISO 27001 does for information security. Proves organization-level governance by audit/certification.
  • Adoption (objective): being a late-2023 standard, it's still early. Few certified organizations (the first US certification was mid-2025), not yet universal like ISO 27001. Not legally required. The adoption driver is customer/procurement trust demand more than regulation. By analogy, it sits roughly where ISO 27001 was 15–20 years ago — a differentiator, not yet table stakes.
  • Original: ISO/IEC 42001:2023 (catalog) — paid, copyrighted (cannot be attached)

IV. Threat / security checklists (hands-on)​

OWASP Top 10 for LLM Applications (2025)​

  • By: OWASP Gen AI Security Project. Free, open license.
  • Content: the 10 major risks of LLM apps (LLM01 prompt injection ~ LLM10 unbounded consumption). Detail in 03. Threat model.
  • Position: the checklist most hands-on in actual security assessment. Unlike abstract frameworks, it becomes assessment items directly.
  • Original: OWASP LLM Top 10 (official) — free PDF download

MITRE ATLAS​

  • By: MITRE. Free.
  • Content: attack tactics/techniques against AI systems in an ATT&CK-style matrix (case-based).
  • Position: reference for "real attack scenarios" during threat modeling/red-teaming.
  • Original: MITRE ATLAS (official)

V. Secure-development frameworks​

Google SAIF (Secure AI Framework)​

  • A practical framework presenting six core elements of AI-system security (governance/detection/response views).
  • Original: Google SAIF (official)

CISA · NCSC "Guidelines for Secure AI System Development"​

  • Joint US CISA / UK NCSC. Security principles across the full design–development–deployment–operation lifecycle.
  • Original: CISA guide (official)

VI. Law / regulation (mandatory)​

EU AI Act​

  • By/nature: EU, the world's first comprehensive AI law. Risk-tiered (prohibited/high/limited/minimal). Extraterritorial (affects non-EU firms serving the EU).
  • Timeline (2025–2026): 2025-02 prohibited practices → 2025-08 general-purpose AI (GPAI) obligations → 2026-08 general application (Art. 50 transparency) → 2026-12 synthetic-output watermarking, etc. (the AI Omnibus adjusted part of the high-risk schedule)
  • Original: EU AI Act portal · official text, EUR-Lex (Regulation (EU) 2024/1689)

Korea — AI Framework Act​

  • Formal name: "Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation of Trust." In force since 22 January 2026.
  • Key: safety obligations for high-impact AI, output marking (watermark), a domestic representative for overseas businesses. Regulated parties are developers/service providers (not ordinary users), with a grace period of a year or more.
  • Original: National Law Information Center (law.go.kr) — search by the law's name

Korea — PIPA + PIPC guide​

VII. Collected originals​

StandardNatureFreeOriginal
NIST AI RMF (100-1)Framework✅link
NIST GenAI Profile (600-1)Framework✅link
ISO/IEC 42001Certification❌ paidcatalog
ISO/IEC 23894Guidance❌ paidcatalog
OWASP LLM Top 10Checklist✅link
MITRE ATLASChecklist✅link
Google SAIFSecure dev✅link
CISA/NCSC guideSecure dev✅link
EU AI ActLaw✅EUR-Lex
AI Framework ActLaw (KR)✅law.go.kr
PIPA / guideLaw (KR)✅pipc.go.kr

:::caution ISO copyright note ISO/IEC 42001 and 23894 are paid, copyrighted documents. Only the official purchase (catalog) links are provided; their original PDFs cannot be attached/redistributed on this site. The others (NIST, OWASP, EU, Korean law) are officially free and downloadable from the links above. :::

VIII. Conclusion​

  • There is no single main standard. Law (mandatory) · certification (ISO 42001) · framework (NIST) · checklist (OWASP) divide the roles and stack up.
  • Practical depth priority: start with the trio NIST AI RMF (frame) + OWASP LLM Top 10 (assess) + domestic law (mandatory); ISO 42001 when certification is needed; ISO 23894 as reference.
  • Since the standards translate into one another via crosswalks, learning one framework (NIST) well makes it easy to extend to the rest.