Buyers of AI services keep asking vendors for proof of AI security, and most vendors answer with a SOC 2 report or an ISO 27001 certificate that was never scoped to the model, the agent runtime or the training data. The Cloud Security Alliance (CSA) STAR for AI programme is an attempt to close that gap with a public, levelled assurance scheme built on the AI Controls Matrix (AICM) and its questionnaire, the AI-CAIQ. This guide explains the levels and routes, what the AI-CAIQ asks, how the identity and non-human identity (NHI) controls are assessed, and how to plan a submission or read one as a buyer. It is written for security, compliance and third-party risk practitioners at AI vendors and at the organisations that buy from them. It covers the assurance scheme; the control content itself is covered in the CSA AI Controls Matrix Guide.
Key takeaways
- Two levels, six paths. Level 1 is a self-assessment (with an optional automated Valid-AI-ted check); Level 2 is third-party assurance through an ISO/IEC 27001 route, an ISO/IEC 42001 route, a combined 27001 plus 42001 route or a SOC 2 attestation route.
- The AICM is the yardstick for every path. Version 1.1 (released 22 June 2026) has 247 control objectives in 18 domains, and its AI-CAIQ has 320 questions.
- ISO/IEC 42001 is optional, not mandatory. CSA states the ISO/IEC 27001 route extends an existing certification to AI security through the AICM without requiring ISO/IEC 42001.
- Scope is where assurance is won or lost. The AICM assigns controls to five roles (model provider, orchestrated service provider, application provider, AI customer, cloud service provider), so a listing only means something if the vendor's role and system boundary are stated.
- Identity evidence is the weak spot. Agent tokens, model API keys and service accounts in pipelines are where AI services are actually breached, so buyers should test the IAM and logging answers hardest.
What it is and who it applies to
STAR (Security, Trust, Assurance and Risk) has been CSA's cloud assurance programme for years, using the Cloud Controls Matrix (CCM) and the Consensus Assessment Initiative Questionnaire (CAIQ). STAR for AI applies the same model to AI services. CSA describes it as its assurance programme for AI services, using the AICM as the framework and the STAR Registry as the public record, and aimed at organisations that build, deploy or embed AI in their services. That covers foundation model providers, companies that wrap or orchestrate models into products, SaaS vendors adding AI features, and cloud platforms hosting AI workloads.
It is voluntary. No regulator requires a STAR for AI listing. Its value is commercial: CSA positions listings for AI third-party risk management, procurement and vendor due diligence, AI supply chain risk and governance work. For a buyer, a listing replaces part of a bespoke questionnaire; for a vendor, one AI-CAIQ answered well can be reused across many customers.
- Vendors: any provider of an AI model, AI platform or AI-enabled application that customers assess. Smaller vendors usually start at Level 1; vendors that already hold ISO/IEC 27001 or SOC 2 can aim for Level 2 by extending their existing audit.
- Buyers: third-party risk teams, procurement and security architects who need a consistent baseline to compare AI suppliers. See the Third-Party AI Risk Guide.
- Auditors: Level 2 engagements use CSA-approved certification bodies and assessment firms listed in CSA's Certified STAR Auditors directory.
STAR for AI sits beside, not inside, the classic cloud STAR programme. If you already know CCM and STAR, the Cloud Security Alliance CCM and STAR Guide covers the cloud scheme; this page covers what changes for AI.
Structure and key requirements
Levels and routes
CSA publishes the following paths on its STAR for AI pages (checked 27 October 2026).
| Level | Path | Standard behind it | Independent check | Best fit |
|---|---|---|---|---|
| Level 1 | Self-Assessment | AICM via AI-CAIQ | None | First listing, early-stage vendors, transparency for customers |
| Level 1 | Valid-AI-ted | AICM via AI-CAIQ | Automated validation of the AI-CAIQ answers; adds a Valid-AI-ted badge | Vendors who want a quality signal before paying for an audit |
| Level 2 Certification | ISO/IEC 27001 route | ISO/IEC 27001 plus AICM | Third-party audit | Vendors with an existing ISMS certificate (or classic STAR certification) |
| Level 2 Certification | ISO/IEC 27001 plus 42001 route | Both ISO standards plus AICM | Third-party audit | Vendors that need both security and AI management system assurance, often for EU customers |
| Level 2 Certification | ISO/IEC 42001 route | ISO/IEC 42001 plus AICM | Third-party audit | AI-first vendors that built an AI management system before an ISMS |
| Level 2 Attestation | SOC 2 route ("STAR for AI Attestation") | SOC 2 plus AICM | Third-party attestation | US-facing vendors whose customers already read SOC 2 reports |
CSA's certification page states that each Level 2 route is paired with an AI-CAIQ run through Valid-AI-ted, and its STAR for AI page recommends submitting the ISO certificate for the certification routes. CSA's pages do not state validity periods or surveillance cycles for STAR for AI listings; check the current programme documents and your certification body's terms rather than assuming the cloud STAR cycle carries over.
The AI Controls Matrix and AI-CAIQ
- Size: AICM v1.1 has 247 control objectives across 18 domains (up from 243 in v1.0) and is synchronised with CCM v4.1. CSA's v1.1 announcement says the AI-CAIQ has 320 questions.
- Domains: the domain set follows the CCM (for example Identity & Access Management, Logging & Monitoring, Cryptography, Encryption and Key Management, and Supply Chain Management, Transparency and Accountability) and adds a Model Security (MDS) domain with 13 AI-specific controls. Infrastructure and Virtualization Security is renamed Infrastructure Security (I&S).
- Five analysis pillars: each control is tagged by control type, applicability and ownership, architectural relevance, LLM lifecycle relevance and threat category. The ownership pillar is what makes scoping possible.
- Roles: model provider (MP), orchestrated service provider (OSP), application provider (AP), AI customer (AIC) and cloud service provider (CSP), each with its own implementation and auditing guidelines.
- Mappings: CSA publishes mappings with gap analysis to BSI AIC4, ISO/IEC 42001:2023, the EU AI Act, AIUC-1, NIST AI RMF and NIST AI 600-1, plus ISO/IEC 27001.
The practical consequence: a model provider and an application provider answering the same AI-CAIQ will mark different controls as owned, shared or inherited. A buyer who reads "compliant" without checking the ownership column learns very little.
Submission process
CSA describes five stages: prepare, complete the assessment, submit through the STAR Platform, get listed in the STAR Registry, then maintain and advance the listing. CSA released a free STAR for AI Prep Kit on 26 October 2026, alongside a Level 1 submission guide inside the AICM package.
Identity and NHI requirements
AI services are run by machine identities: model API keys, agent access tokens, service accounts in training and inference pipelines, workload identities for GPUs and vector stores, and OAuth grants that let agents act in customer systems. The AICM carries the CCM's Identity & Access Management domain into the AI context, and the AI-CAIQ asks the vendor to describe how those identities are controlled. When you answer or review those questions, push for evidence on the points below.
| Identity area | What a credible AI-CAIQ answer shows | Evidence to request |
|---|---|---|
| Agent and service identities | Each agent, pipeline and tool integration has its own identity and owner; no shared keys across tenants | Identity inventory extract with owners; tenant isolation design |
| Model and API keys | Keys are vaulted, scoped, rotated and revocable; short-lived tokens preferred over static keys | Secrets management configuration; rotation logs; key scope list |
| Customer-facing tokens | Tokens issued to customers' agents are scoped per tenant and per action and can be revoked by the customer | Token lifetime and scope settings; revocation procedure |
| Privileged access to models and data | Access to training data, weights and production inference is least privilege and reviewed | Access review records for model repositories and data stores |
| Logging of non-human activity | Agent and API calls are logged with the identity used, and logs are retained and monitored | Sample log schema; alerting rules for anomalous token use |
This is not theoretical. In the Moltbook breach, a database key shipped in public JavaScript with Row Level Security disabled exposed about 1.5 million agent API tokens, and because each key was the agent's identity, anyone who read the table could impersonate an agent. An AI-CAIQ answer that says "API keys are protected" without describing storage, scope and revocation would not have surfaced that risk. See the Agent Identity and Access Guide and the Credential Vaulting for AI Agents Guide.
Implementation steps
- Decide your role and boundary. State which AICM roles you play (often more than one, for example application provider and orchestrated service provider) and draw the system boundary: models, agents, tools, data stores, hosting.
- Pick the target level. Start at Level 1 if you have no audited ISMS. If you hold ISO/IEC 27001 or SOC 2, plan Level 2 through the matching route so the AI controls ride on an audit you already pay for.
- Download the current AICM, AI-CAIQ and Prep Kit. Work from the published version (v1.1 at the time of writing) and record which version you answered; mixing versions is a common rejection reason in questionnaire programmes.
- Run a gap assessment by owner. For each control mark owned, shared or inherited, then name the internal owner and the evidence. Treat the MDS and IAM domains as priority because they are where AI-specific weaknesses sit.
- Fix the gaps before answering. Typical fixes: an agent and API key inventory, vaulting and rotation for model keys, logging of agent actions with identity context, model change control and red-team testing records.
- Complete the AI-CAIQ honestly. Answer "no" or "partially" with a remediation date where true. Auditors and buyers trust a questionnaire with credible gaps more than one with 320 unqualified "yes" answers.
- Validate and submit. Run the answers through Valid-AI-ted (optional at Level 1, paired with Level 2 routes per CSA), then submit through the STAR Platform for the Registry listing.
- For Level 2, brief your auditor early. Agree with a CSA-approved certification body or assessment firm how AICM controls will be sampled within your ISO or SOC 2 engagement and which evidence they expect.
- Maintain the listing. Update the AI-CAIQ when models, hosting or agent capabilities change materially, and when CSA releases a new AICM version.
Evidence and assessment
What an assessor or buyer can rely on differs sharply by path. Use this table to calibrate trust.
| Path | Who checks the answers | What it proves | What it does not prove |
|---|---|---|---|
| Level 1 Self-Assessment | Nobody independent | The vendor has documented its AI controls publicly and is accountable for the statements | That the controls exist or operate |
| Level 1 Valid-AI-ted | Automated validation by CSA | The answers have passed CSA's automated quality check | Operating effectiveness; CSA's pages do not describe the method in detail |
| Level 2 ISO routes | CSA-approved certification body | A management system audit covered AI controls within a defined scope | Anything outside the certificate scope |
| Level 2 SOC 2 route | Independent service auditor | Controls were tested against criteria, with any exceptions reported | Controls outside the report's system description or period |
For a buyer the minimum review is: check the Registry entry date and AICM version; confirm the listed service is the one you are buying; read the scope and role; read every "no" and "partial" answer in IAM, LOG, MDS and STA; and for Level 2, obtain the certificate or SOC 2 report and confirm the AI system is in the stated scope. The How to Read a SOC 2 Report Guide covers the report side.
For a vendor, keep an evidence register keyed by AICM control ID: policy reference, system configuration export, log sample, owner and last test date. That register is what makes a Level 2 audit quick and what lets you answer customer follow-up questions without re-collecting evidence.
Mapping to other frameworks
STAR for AI is an assurance wrapper; the mapping work below shows where its evidence can be reused. Mappings to ISO/IEC 42001, the EU AI Act and NIST AI RMF come from CSA's AICM package; the NIST SP 800-53 and CSF entries are our own alignment.
| Framework | Relationship to STAR for AI | Where evidence is reused |
|---|---|---|
| ISO/IEC 42001:2023 | Certification route; AICM mapped with gap analysis | AI risk assessment, AI impact assessment, AI system life cycle records |
| ISO/IEC 27001:2022 | Certification route; AICM mapped | Annex A access control, supplier and logging controls extended to AI systems |
| SOC 2 | Attestation route | Security and availability criteria testing for the AI service |
| EU AI Act | AICM mapped with gap analysis | Technical documentation, logging and risk management evidence for providers and deployers. See the EU AI Act Compliance Guide. |
| NIST AI RMF and AI 600-1 | AICM mapped | Generative AI risk controls. See the NIST AI 600-1 Generative AI Profile Guide. |
| AIUC-1 | AICM mapped (new in v1.1) | Agent-specific controls. See the AIUC-1 Agent Certification Guide. |
| NIST SP 800-53 Rev. 5 | Buyer-side use of listings | SR-6 Supplier Assessments and Reviews, SA-9 External System Services, CA-2 Control Assessments; IAM answers align to IA-5 Authenticator Management, IA-9 Service Identification and Authentication, AC-6 Least Privilege |
| NIST CSF 2.0 | Buyer-side use of listings | GV.SC-06 (due diligence before entering supplier relationships), GV.SC-07 (supplier risks understood and monitored) |
For a broader view of how AI frameworks overlap, see the AI Security Frameworks Crosswalk Guide.
Common pitfalls
- Treating Level 1 as certification. A self-assessment badge is a disclosure, not an audit. Buyers who accept it as equivalent to Level 2 skip the evidence step entirely.
- Scope that excludes the AI. An ISO/IEC 27001 certificate scoped to "corporate IT" or a SOC 2 report whose system description omits the model platform does not cover the AI service, even if the vendor lists it under STAR for AI routes. Check the scope statement.
- Ignoring the ownership column. Inherited controls from a cloud or model provider still need the upstream provider's evidence. Ask which provider and request their listing too.
- Static API keys described as "secure". Long-lived model and agent keys with broad scope are the most common real-world failure, as Moltbook showed. Ask for lifetimes, scopes and revocation tests.
- Stale listings. AI services change monthly. A listing answered against AICM v1.0 for a product that has since added autonomous agents tells you about a different system.
- Answering for marketing. Vendors who answer "yes" everywhere lose credibility with experienced reviewers and create contractual exposure if a breach later shows the statement was untrue.
Practitioner checklist
- Record your AICM role(s) and draw the AI system boundary before opening the AI-CAIQ.
- Download the current AICM, AI-CAIQ and Prep Kit, and note the version in every artefact.
- Run a gap assessment by control owner, prioritising the IAM, LOG, MDS and STA domains.
- Build an inventory of agent identities, model API keys and service accounts with owners, scopes and rotation dates.
- Choose the Level 2 route that matches the ISO or SOC 2 audit you already hold, and confirm the AI service is inside its scope.
- Use a CSA-approved certification body or assessment firm and agree AICM sampling up front.
- As a buyer, read every "no" and "partial" answer and request evidence for identity and logging controls.
- Request upstream providers' listings for every inherited control.
- Re-answer the AI-CAIQ when the AICM version, hosting or agent capabilities change.
Standards and references
- Cloud Security Alliance: STAR for AI
- Cloud Security Alliance: About STAR for AI
- Cloud Security Alliance: AI Controls Matrix v1.1
- Cloud Security Alliance: AI Controls Matrix v1.1, Strengthening the Foundation for Trustworthy AI
- Cloud Security Alliance: STAR Certification
- Cloud Security Alliance: STAR for AI Prep Kit
- Cloud Security Alliance: Cloud Controls Matrix
- ISO: ISO/IEC 42001:2023 Artificial intelligence management system
Related NHI Mgmt Group resources: CSA AI Controls Matrix Guide · Cloud Security Alliance CCM and STAR Guide · ISO/IEC 42001 Guide · AI Vendor Security Questionnaire Template · Third-Party AI Risk Guide