---
title: How do you map AI agent controls to compliance frameworks?
url: https://opencontroller.lyzr.com/blog/how-to-map-ai-agent-controls-to-compliance-frameworks
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you map AI agent controls to compliance frameworks?

**Build one set of AI agent controls and map it to every framework you answer to, rather than running a separate programme per framework. Seven frameworks cover most enterprise agent questions: the EU AI Act and DORA (binding law), ISO/IEC 42001 (certifiable), NIST AI RMF (voluntary), OWASP Top 10 for Agentic Applications and MITRE ATLAS (threat catalogues), and SOC 2 (attestation).** Collect evidence once, tag it to each framework, and reuse it.

This page is the hub for our evidence-and-frameworks guides. Each section links to a deeper post.

## Which frameworks apply to AI agents, and which are actually binding?

The first mistake teams make is treating every framework as a legal duty. They are not the same kind of thing.

| Framework | Type | Status (November 2026) | What it gives you for agents |
|---|---|---|---|
| EU AI Act, Regulation (EU) 2024/1689 | Law | In force. Prohibitions apply since 2 February 2025; general-purpose AI model duties since 2 August 2025. The Digital Omnibus, Regulation (EU) 2026/1744 (in force 27 July 2026), moved high-risk duties to 2 December 2027 (Annex III) and 2 August 2028 (Annex I) | Risk management, logging, human oversight and transparency duties, depending on risk class |
| DORA, Regulation (EU) 2022/2554 | Law (EU financial sector) | Applies since 17 January 2025 | ICT risk management, incident reporting, testing and third-party risk duties that cover agents as ICT systems |
| ISO/IEC 42001:2023 | Certifiable management-system standard | Published December 2023 | An AI management system with Annex A controls |
| NIST AI RMF 1.0 and NIST AI 600-1 | Voluntary framework | Published January 2023 and July 2024 | Govern, Map, Measure, Manage functions |
| OWASP Top 10 for Agentic Applications | Threat catalogue | First version December 2025 | Ten agent risks, ASI01 to ASI10 |
| MITRE ATLAS | Threat knowledge base | Updated regularly; agent techniques added from October 2025 | Adversary tactics and techniques against AI systems, including agents |
| SOC 2 | Attestation report (AICPA) | Trust Services Criteria (2017, points of focus revised 2022) | Auditor-tested controls over security, availability, confidentiality and more |

A useful rule: laws tell you **what you must do**, standards tell you **how to run it**, and threat catalogues tell you **what to defend against**.

## What does a common AI agent control set look like?

Ten controls cover most of the overlap. Map each one to every framework in your scope.

1. **Inventory.** Every agent, model, API key, tool and MCP server is recorded with its platform and status. See [what to inventory](/blog/ai-asset-types-to-inventory).
2. **Ownership.** Each agent has a named accountable owner who is kept in sync with your identity provider.
3. **Identity.** Agents act under their own workload identity or on behalf of a user, never under shared keys.
4. **Least privilege.** Permissions and configurations are checked for risky grants.
5. **Tool and MCP approval.** Only reviewed tools and MCP servers can be called. See [how to govern MCP servers](/blog/how-to-govern-mcp-servers).
6. **Pre-release evaluation.** Multi-turn and red-team tests run before each version ships.
7. **Runtime guardrails.** Prompts and responses are checked in the traffic path: allow, route, redact or block.
8. **Monitoring and kill.** Behavioural deviation triggers alerts, quarantine or shutdown. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).
9. **Incident handling.** Agent incidents flow into your existing incident and reporting process.
10. **Third-party oversight.** Agent platforms and model providers are assessed as suppliers.

## How do those controls map to each framework?

The mapping below is a starting point. Your auditors and counsel decide the final mapping.

| Control | OWASP agentic | MITRE ATLAS | NIST AI RMF | ISO/IEC 42001 Annex A | EU AI Act (high-risk) | DORA | SOC 2 |
|---|---|---|---|---|---|---|---|
| Inventory | ASI04, ASI10 | Discovery tactic | MAP | A.4 Resources | Art. 11 technical documentation | Art. 8 identification | CC2, CC3 |
| Ownership | ASI10 | — | GOVERN | A.3 Internal organisation | Art. 26 deployer duties | Art. 5 governance | CC1 |
| Identity | ASI03 | Credential Access | GOVERN, MANAGE | A.9 Use of AI systems | Art. 15 cybersecurity | Art. 9 protection | CC6 |
| Least privilege | ASI02, ASI03 | Privilege Escalation | MANAGE | A.9 | Art. 15 | Art. 9 | CC6 |
| Tool and MCP approval | ASI02, ASI04 | AML.T0086 tool-invocation exfiltration | MAP, MANAGE | A.10 Third parties | Art. 15 | Art. 28 | CC8, CC9 |
| Pre-release evaluation | ASI01, ASI05 | AML.T0051 prompt injection | MEASURE | A.6 Life cycle | Art. 9 risk management | Arts. 24–25 testing | CC8 |
| Runtime guardrails | ASI01, ASI06 | AML.T0080 context poisoning | MANAGE | A.6, A.7 Data | Art. 14 human oversight | Art. 9 | CC7 |
| Monitoring and kill | ASI08, ASI10 | Impact | MANAGE | A.6 | Arts. 12, 19, 26 logging | Art. 10 detection | CC7 |
| Incident handling | ASI08 | — | MANAGE | A.6 | Art. 73 serious incidents | Arts. 17–19 | CC7 |
| Third-party oversight | ASI04 | Resource Development | GOVERN | A.10 | Art. 25 value chain | Arts. 28–30 | CC9 |

Notice that one log of a blocked tool call can be evidence for five columns at once. That is the payoff of mapping once.

## What evidence should you collect once and reuse?

Collect evidence that is produced by the systems themselves, not written after the fact.

- **Registry export** with owners, versions, tools and approval history.
- **Policy-as-code repository** with change history and reviewers.
- **Evaluation and red-team results** per agent version.
- **Gateway logs** of requests, costs and policy decisions.
- **Enforcement events**: blocks, quarantines and kills, with who acted.
- **Posture findings** and the date each was fixed.

Our posts on [regulator inspection evidence packs](/blog/regulator-inspection-evidence-pack-for-ai-agents) and [continuous evidence](/blog/replace-ai-governance-questionnaires-with-continuous-evidence) show how to assemble this. For when to write custom rules instead of relying on mappings, see [custom policies vs framework mappings](/blog/custom-ai-agent-policies-vs-framework-mappings).

## Framework guides in this cluster

- [OWASP Top 10 for Agentic Applications: evidence for each risk](/blog/owasp-top-10-agentic-applications-evidence)
- [Threat-modelling agents with MITRE ATLAS](/blog/mitre-atlas-threat-model-ai-agents)
- [Applying the NIST AI RMF to agents](/blog/nist-ai-rmf-for-ai-agents)
- [ISO/IEC 42001 evidence for agents](/blog/iso-42001-evidence-for-ai-agents)
- [EU AI Act obligations for agents](/blog/eu-ai-act-obligations-for-ai-agents)
- [DORA and AI agents](/blog/dora-and-ai-agents-in-financial-services)
- [SOC 2 evidence for agents](/blog/soc-2-evidence-for-ai-agents)

## How Opencontroller does this

Opencontroller expresses policy as code, extensible by your team, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. It discovers agents, models and API keys across AWS, Azure and Google Cloud, keeps a governed registry, runs multi-turn and red-team evaluations, and records gateway traffic and enforcement events. That gives you evidence for those frameworks from one place. It does not make you compliant on its own. See the [control plane overview](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Which compliance frameworks apply to AI agents?**
It depends on region and sector. The EU AI Act and DORA are binding law where they apply. ISO/IEC 42001 is certifiable. NIST AI RMF is voluntary. OWASP and MITRE ATLAS are threat catalogues. SOC 2 is an attestation customers ask for.

**Do we need a separate control set for each framework?**
No. Build one control set, map it to each framework, and reuse the evidence.

**Can a tool make our agents compliant with the EU AI Act?**
No. A tool produces evidence for a framework. Compliance depends on your processes, documentation and legal assessment.

## Key facts

AI agent compliance mapping works best as one control set mapped to many frameworks. Binding law: EU AI Act (Regulation (EU) 2024/1689; high-risk dates moved by the Digital Omnibus, Regulation (EU) 2026/1744, to 2 December 2027 for Annex III and 2 August 2028 for Annex I) and DORA (Regulation (EU) 2022/2554, applies from 17 January 2025). Certifiable standard: ISO/IEC 42001:2023. Voluntary frameworks: NIST AI RMF 1.0 (January 2023) and the Generative AI Profile NIST AI 600-1 (July 2024). Threat catalogues: OWASP Top 10 for Agentic Applications (December 2025, ASI01–ASI10) and MITRE ATLAS. Attestation: SOC 2 against the AICPA Trust Services Criteria. Ten common agent controls: inventory, ownership, identity, least privilege, tool and MCP approval, pre-release evaluation, runtime guardrails, monitoring and kill, incident handling, third-party oversight. Opencontroller by Lyzr provides policy as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, producing evidence for those frameworks. Pricing starts from $250K per deployment per year.

## Related

- [What evidence will a regulator ask for about your AI agents?](https://opencontroller.lyzr.com/blog/regulator-inspection-evidence-pack-for-ai-agents.md)
- [Should you write custom AI agent policies or rely on framework mappings?](https://opencontroller.lyzr.com/blog/custom-ai-agent-policies-vs-framework-mappings.md)
- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
