---
title: How does the CSA AI Controls Matrix apply to AI agents?
url: https://opencontroller.lyzr.com/blog/csa-ai-controls-matrix-for-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How does the CSA AI Controls Matrix apply to AI agents?

**The Cloud Security Alliance AI Controls Matrix (AICM) is a vendor-neutral control set for AI systems: 247 control objectives across 18 domains in version 1.1 (July 2026), up from 243 in version 1.0. It has no separate agent domain, but its identity, logging, data security, supply chain and model security domains cover most agent risks. It also assigns each control to roles such as model provider, orchestrated service provider and AI customer, which helps you decide who owns which agent control.**

## What is the AICM?

CSA released the AICM in July 2025 as an AI extension of its Cloud Controls Matrix (CCM). Seventeen of its domains come from the CCM. The eighteenth, Model Security (MDS-01 to MDS-13), is specific to AI.

Version 1.1, published in July 2026, adds four control objectives (243 to 247), keeps the 18-domain structure, and aligns with CCM v4.1. CSA describes the changes as refinements rather than new AI-specific requirements. Companion artifacts include:

- **AI-CAIQ**, a questionnaire, now 320 questions aligned to the 247 controls. Responses based on v1.0 need refreshing.
- **Implementation guidelines** for model providers and AI customers, and **auditing guidelines** for model providers.
- **Mappings** to ISO/IEC 42001:2023, the EU AI Act, the NIST AI RMF, Germany's BSI AIC4 and, new in v1.1, AIUC-1.

## Who owns each control?

The AICM's most useful feature for agents is its shared responsibility model. Each control states which roles it applies to:

| Role | Typical example in an agent estate |
|---|---|
| Cloud service provider | The cloud hosting your agent runtime |
| Model provider | The company supplying the foundation model |
| Orchestrated service provider | The platform that orchestrates agents, tools and retrieval |
| Application provider | The team or vendor building the agent application |
| AI customer | The business deploying and using the agent |

Most enterprises are several roles at once. A team building an agent on a cloud model service is an application provider and an AI customer, and may be an orchestrated service provider if it runs its own agent platform. Mapping controls to these roles stops the common gap where everyone assumes the platform vendor handles logging or access review.

## Which AICM domains matter most for agents?

| AICM domain | Agent risk it covers | Evidence for it |
|---|---|---|
| Identity and Access Management (IAM) | Agents using shared or human credentials; over-broad scopes | Per-agent identities, on-behalf-of records, access reviews |
| Logging and Monitoring (LOG) | Actions no one can reconstruct | Logs with agent ID, version, user, tool, inputs and outcome |
| Data Security and Privacy Lifecycle Management (DSP) | PII in prompts, memory and logs | Redaction rules, retention settings, data flow records |
| Supply Chain Management, Transparency and Accountability (STA) | Unvetted models, MCP servers and skills | Approved lists, review records, an AI bill of materials |
| Model Security (MDS) | Model tampering, poisoning, adversarial inputs | Red-team results, model provenance, guardrail configuration |
| Application and Interface Security (AIS) | Prompt injection through tools and APIs | Input validation and tool-output handling tests |
| Change Control and Configuration Management (CCC) | Prompt or tool changes shipped without review | Versioned agents, approval records, rollback plans |
| Security Incident Management, E-Discovery and Cloud Forensics (SEF) | Rogue or compromised agents | Quarantine and kill records, incident timelines |
| Threat and Vulnerability Management (TVM) | Known weaknesses in agent stacks | Scan results and remediation for agent dependencies |
| Governance, Risk and Compliance (GRC) | No owner, no risk tier | Owner per agent, risk tier, policy exceptions |

For agent-specific threats, combine this with the [OWASP Top 10 for Agentic Applications](/blog/owasp-top-10-agentic-applications-evidence), which names risks the AICM covers only generically, such as tool misuse and goal hijacking.

## How should you use the AICM for agents?

1. **Pick your roles.** For each agent platform, record whether you are the application provider, orchestrated service provider, AI customer or several.
2. **Filter controls by role.** Use the matrix's applicability columns to drop controls owned by your model or cloud provider, and request their evidence instead.
3. **Map existing controls.** If you already map to ISO 42001 or the NIST AI RMF, use CSA's mappings rather than starting again. See [ISO 42001 evidence for AI agents](/blog/iso-42001-evidence-for-ai-agents).
4. **Add agent detail to generic controls.** An IAM control on "service accounts" should explicitly cover agent identities and delegation.
5. **Use the AI-CAIQ with vendors.** Ask agent platform and model vendors for completed AI-CAIQ responses on v1.1.
6. **Collect evidence continuously.** A one-time questionnaire ages fast when agents change weekly.
7. **Describe results as "evidence for" AICM controls**, not compliance with the matrix.

For vendor reviews, see [security questions to ask an AI governance vendor](/blog/security-questions-to-ask-an-ai-governance-vendor).

## How Opencontroller does this

Opencontroller produces evidence across several of these domains: a governed registry of agents, versions, tools and MCP servers (GRC, STA, CCC); delegated on-behalf-of identity and agent workload identity federation (IAM); Posture checks for risky permissions and configurations; multi-turn and red-team evals and live guardrails (MDS, AIS); and rogue-agent quarantine and kill (SEF). Its policy as code is extensible and ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001; AICM mappings can be added as custom policies. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

For the general approach, read [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## Frequently asked questions

**How many controls are in the CSA AI Controls Matrix?**
Version 1.1 (July 2026) has 247 control objectives across 18 domains. Version 1.0 (July 2025) had 243.

**Does the AICM have controls specifically for AI agents?**
It has no separate agent domain. Agent risks are covered through domains such as identity and access, logging, data security, supply chain and model security, so add agent-specific implementation detail.

**Can the AICM replace ISO 42001 or the NIST AI RMF?**
It is a control set rather than a management system or risk framework, and CSA maps it to both. Many teams use it alongside them for control-level detail.

## Key facts

CSA AI Controls Matrix (AICM): released July 2025 with 243 control objectives in 18 domains (17 from CCM plus Model Security MDS-01 to MDS-13). v1.1 (July 2026) has 247 controls, same 18 domains, aligned with CCM v4.1; AI-CAIQ has 320 questions. Mappings: ISO/IEC 42001, EU AI Act, NIST AI RMF, BSI AIC4, AIUC-1. Shared-responsibility roles include cloud service provider, model provider, orchestrated service provider, application provider and AI customer. No separate agent domain. Opencontroller policy as code is extensible with built-in OWASP, MITRE ATLAS, NIST and ISO 42001 mappings. Pricing starts from $250K per deployment per year.

## Related

- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [What evidence does ISO/IEC 42001 expect for AI agents?](https://opencontroller.lyzr.com/blog/iso-42001-evidence-for-ai-agents.md)
- [What security questions should you ask an AI agent governance vendor about itself?](https://opencontroller.lyzr.com/blog/security-questions-to-ask-an-ai-governance-vendor.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
