Learn
ISO 42001 and the NIST AI RMF for AI agents.
Both frameworks ask you to prove that your AI controls work. For agents, that proof comes from runtime records.
Updated 31 October 2026. This page is not legal or certification advice.
How do ISO 42001 and the NIST AI RMF apply to AI agents?
ISO/IEC 42001:2023 is a certifiable management-system standard for AI: policies, roles, risk and impact assessment, and controls listed in Annex A. The NIST AI Risk Management Framework is voluntary guidance organised in four functions: Govern, Map, Measure and Manage. Neither names agents, but both cover any AI system you run, and an agent that calls tools with delegated authority is exactly where controls must be shown to operate, not just written down.
The published NIST framework is still AI RMF 1.0 (NIST AI 100-1, January 2023), with the Generative AI Profile (NIST AI 600-1) from July 2024. NIST says a revision is under way; no revised version had been published at the time of writing.
Which runtime evidence fits Govern, Map, Measure and Manage?
Govern is cross-cutting; the other three run through the life of a system.
| NIST AI RMF function | What it asks, for an agent | Runtime evidence MAQPNA produces | Not covered by MAQPNA |
|---|---|---|---|
| Govern | Policies, accountability and roles; who may approve what; third-party risk | Policies as versioned code (ToolPolicy), separation of duties for approvers, a third-party register of MCP servers, model routes and agent peers used | Writing the AI policy, board accountability, training and culture |
| Map | Context, intended use, which systems and data an agent touches | AI-BOM (CycloneDX or SPDX) of agents, images, models, MCP servers, tool pins and memory stores | Use-case and impact analysis, stakeholder input |
| Measure | Track risks and whether controls work | Denials by reason, DLP blocks, approvals, budget stops from the ledger; OpenTelemetry GenAI traces; maqpna doctor posture score | Model accuracy, bias, robustness and red-team evaluation |
| Manage | Treat, respond to and recover from risk | Default-deny tool policy, approvals, the kill switch, revocations and a verified incident timeline | Incident communication, decisions to retire a system |
Which ISO 42001 Annex A areas can runtime evidence support?
Annex A has nine control objectives, A.2 to A.10. Area names are summarised; check control wording in your licensed copy of the standard.
| Annex A area | Runtime evidence MAQPNA produces | Coverage |
|---|---|---|
| A.2 Policies related to AI | Tool policies and sovereignty policies as code, with versions | Partial: the written AI policy is yours |
| A.3 Internal organisation | Approver roles and groups, no self-approval, approver identity in each record | Partial |
| A.4 Resources for AI systems | Inventory of agents, models, tools and sandbox tiers | Partial |
| A.5 Assessing impacts of AI systems | None directly; the ledger can inform an assessment | Not covered by MAQPNA (customer process) |
| A.6 AI system life cycle, including A.6.2.8 recording of event logs | Hash-chained audit ledger with signed checkpoints; policy test and replay before rollout | Supports evidence for the operation and logging controls |
| A.7 Data for AI systems | DLP detector counts, taint labels, residency and egress allow-lists | Partial: training-data provenance is out of scope |
| A.8 Information for interested parties | Signed evidence packs an auditor can verify offline | Partial: user-facing information is yours |
| A.9 Use of AI systems | Per-tool policy, approvals, budgets, kill switch, session timeline | Supports evidence for responsible use |
| A.10 Third-party and customer relationships | Third-party register, tool pinning, registry allow-lists | Partial: contracts and supplier review are yours |
What are Forrester AEGIS and the NIST AI Agent Standards Initiative?
Forrester AEGIS (Agentic AI Enterprise Guardrails for Information Security) is an analyst framework Forrester introduced in August 2025. It has six domains: governance, risk and compliance; identity and access management; data security and privacy; application security; threat management; and Zero Trust architecture. Its principles include least agency, continuous assurance and explainable outcomes. The full report is for Forrester clients, so we do not map it control by control here.
The NIST AI Agent Standards Initiative was announced on 17 February 2026 by NIST's Center for AI Standards and Innovation (CAISI). It has three pillars: industry-led agent standards, open-source agent protocols, and research on agent security and identity. It began with a request for information on AI agent security and a concept paper on agent identity and authorisation. NIST has said further guidelines will follow; treat anything beyond that as provisional until NIST publishes it.
Both point the same way as this page: distinct agent identities, least agency, and continuous, checkable evidence. Those are what AI agent governance at runtime supplies.
What do India's DPDP Act and RBI FREE-AI mean for AI agents?
India's Digital Personal Data Protection Act, 2023 applies when agents process personal data. The DPDP Rules, 2025 were notified in November 2025 with an 18-month phased timeline, so most obligations on data fiduciaries apply from around May 2027. They include reasonable security safeguards, breach notification and erasure. MAQPNA's dpdp profile warns below 365 days of log retention, and maqpna memory erase deletes a person's entries from governed memory stores with a signed certificate. It does not reach data inside upstream tools, sandboxes or prompts already sent to model providers.
The Reserve Bank of India's FREE-AI committee report (13 August 2025) sets out seven principles ("Sutras") and 26 recommendations for AI in the financial sector, including board-approved AI policies, audits and attention to third-party AI providers. It is a committee framework; check for any directions RBI has issued since. Ledger evidence and the third-party register support the audit and outsourcing parts. For CERT-In's 180-day log retention and 6-hour reporting, see the compliance profiles.
Where MAQPNA fits
Evidence for the management system, not the certificate.
MAQPNA runs your agents behind one gateway that enforces policy, holds risky calls for a person, can stop an agent at once and writes a tamper-evident audit trail. An evidence pack for iso42001 turns that trail into a signed report for one system and period, marking each control evidenced, partial or "not covered by MAQPNA (customer process)". Its mapping stays draft until counsel and audit partners review it.
Certification is done by an accredited body against your whole management system. MAQPNA cannot make you certified and Kosa8 Labs holds no certifications. See also the EU AI Act for AI agents, the OWASP agentic top 10 mapping, air-gapped governance, the trust centre and solutions by sector.
FAQ
Common questions
Does ISO 42001 cover AI agents?
Yes, as AI systems in scope of your management system. It does not name agents, but its life-cycle, event-logging, use and third-party controls apply directly to agents that call tools.
What are the four functions of the NIST AI RMF?
Govern, Map, Measure and Manage. Govern is cross-cutting; Map sets context, Measure tracks risks and controls, and Manage treats and responds to them.
Is there a NIST standard for AI agents?
Not yet a final one. NIST's AI Agent Standards Initiative, announced in February 2026, is gathering input on agent security, identity and interoperability, and has said guidelines will follow.
Can MAQPNA get us ISO 42001 certified?
No. Only an accredited certification body can certify your management system. MAQPNA supplies runtime controls and signed evidence that some Annex A controls operated.
Turn your AI policy into runtime evidence.
Govern an agent on your laptop today, then generate a signed ISO/IEC 42001 evidence pack from the ledger.
curl -fsSL https://maqpna.com/install.sh | shirm https://maqpna.com/install.ps1 | iex