Standard vendor security questionnaires ask about encryption, SOC 2 reports and patching, and say nothing about whether a supplier trains on your prompts, what its agents can do with your tokens or how it would detect a prompt injection. This template is a ready-to-send security questionnaire for suppliers of AI products: SaaS with embedded assistants, model APIs, agent platforms and AI features in existing tools. Each question has an ID, the evidence to request and a mapping to the domains and threat categories of the Cloud Security Alliance AI Controls Matrix (AICM) and the six AIUC-1 risk domains, so answers can be compared with a supplier's AI-CAIQ or AIUC-1 report. It sits after business intake: use the AI Procurement Checklist to decide whether to proceed, then send this questionnaire to the vendor.
How to use this template
Customise: replace [Organisation], [Vendor] and other placeholders; delete sections that do not apply (for example section 6 for a supplier with no agent features).
Tier the depth: send sections 3, 4 and 9 to every AI supplier; add sections 5 to 8 for suppliers that process confidential or personal data, act on your systems, or support a high-risk use.
Accept existing evidence: if the vendor has a completed CSA AI-CAIQ (published in the STAR Registry or shared under NDA) or an AIUC-1 certificate and report, map it to these questions and ask only what remains.
Evidence over assertion: each question names the evidence expected. "Yes" without evidence scores as partial.
Record answers: keep responses, evidence and the reviewer's score in the third-party risk record, and re-send on renewal or material change.
1. Scope and triggers
[Organisation] sends this questionnaire when any of the following applies:
A new supplier will provide an AI model, AI-enabled SaaS feature, AI agent or agent platform.
An existing supplier enables a generative AI or agent feature in a product already in use.
A supplier changes its underlying model provider, training-data terms, sub-processors or hosting region.
Annual review of suppliers rated [High] or [Critical] for AI risk.
Response time: [15 working days]. Confidentiality: responses are held under [NDA reference] and used only for supplier risk assessment.
2. Roles and contacts
Role
Responsibility
Contact
Third-party risk lead
Sends the questionnaire, chases responses, owns the risk record
Scores section 4 and decides on DPIA and transfer requirements
[Name]
Business owner
Accepts residual risk and conditions
[Name]
Vendor respondent
Named security contact at the vendor who signs off answers
[Vendor contact]
3. Vendor and service profile
ID
Question
Evidence
AICM domain or threat / AIUC-1 domain
AV-3.1
Describe the AI functionality in scope, the models used (own, third-party, open-weight) and the model providers behind them.
Architecture diagram; model and provider list
Supply Chain Management, Transparency, and Accountability / E
AV-3.2
Which shared-responsibility role(s) do you hold: cloud service provider, model provider, orchestrated service provider, application provider?
Shared responsibility matrix
AICM applicability and ownership pillar / E
AV-3.3
Under the EU AI Act, are you the provider of the system as we will use it, and is any intended use high-risk?
Classification statement; instructions for use
Loss of Governance/Compliance threat / E
AV-3.4
Do you have a completed AI-CAIQ, STAR for AI listing, AIUC-1 certificate, ISO/IEC 42001 certificate or ISO/IEC 27001 certificate covering this service?
Certificates, scope statements, reports
All / All
4. Data protection and training use
ID
Question
Evidence
AICM domain or threat / AIUC-1 domain
AV-4.1
Are our prompts, files, outputs or telemetry used to train or improve any model, yours or a third party's? Is this off by default and contractually excluded?
Contract clause; admin setting screenshot
Data Security and Privacy Lifecycle Management / A
AV-4.2
Where are prompts, outputs, embeddings, logs and fine-tuning data stored and processed, and for how long? Can we set retention to [zero or N days]?
Data flow diagram; retention schedule
Data Security and Privacy Lifecycle Management / A
AV-4.3
List sub-processors that receive our data, including model API providers and observability tools.
Sub-processor list with locations
Supply Chain Management, Transparency, and Accountability / A
AV-4.4
How is our data isolated from other customers in vector stores, caches, session state and fine-tuned models?
Tenant isolation design; test results
Data Security and Privacy Lifecycle Management; Model Security / A
AV-4.5
How do you honour deletion requests across derived stores (embeddings, logs, backups, adapters)?
Deletion procedure; evidence of a completed request
Data Security and Privacy Lifecycle Management / A
AV-4.6
Which controls detect and prevent personal data, secrets and our classified data appearing in outputs to other users?
Output filtering description; test results
Sensitive Data Disclosure threat / A
5. Identity, access and agent permissions
ID
Question
Evidence
AICM domain or threat / AIUC-1 domain
AV-5.1
Do you support SSO (SAML or OIDC), MFA and SCIM provisioning for our users and administrators?
Configuration documentation
Identity and Access Management / B
AV-5.2
When the product acts on our systems, which credentials does it use (OAuth grants, API keys, service accounts), with what scopes, and how long do they live?
Scope list; token lifetime settings
Identity and Access Management / B
AV-5.3
Does each agent or integration have its own identity, or do agents share a platform credential across customers or tasks?
Identity architecture
Identity and Access Management / B
AV-5.4
Can we restrict which tools, connectors and MCP servers an agent may use, and require human approval for specific actions?
Admin policy screenshots
Insecure Apps/Plugins threat / B, C
AV-5.5
How are our credentials stored at your side, who among your staff can access them, and is that access logged?
What controls address direct and indirect prompt injection, and how do you limit what an injected instruction can do (tool restrictions, output channel controls)?
Design description; latest test results
Model Manipulation threat / B
AV-6.2
Do you block data exfiltration through rendered links and images in model output?
Configuration; test case
Sensitive Data Disclosure threat / A, B
AV-6.3
How often is the AI system red-teamed or adversarially tested, by whom, and against which framework (for example OWASP Top 10 for LLM Applications, MITRE ATLAS)?
Executive summary of the latest test; remediation status
Model Security / B, C
AV-6.4
How do you protect model weights, system prompts and configuration from theft or unauthorised change?
Access controls; change records
Model Theft threat / B
AV-6.5
What rate limits, cost controls and abuse detection protect the service from denial of service or runaway agent usage?
Limits documentation
Denial of Service threat / D
AV-6.6
How do you notify us before changing the model, model version or safety settings, and can we pin a version?
Change notification policy
Loss of Governance/Compliance threat / D, E
7. AI supply chain
ID
Question
Evidence
AICM domain or threat / AIUC-1 domain
AV-7.1
Do you maintain an AI bill of materials (models, datasets, libraries, plugins) for the service?
AI-BOM or SBOM extract
Supply Chain Management, Transparency, and Accountability / E
AV-7.2
How do you verify the integrity and provenance of third-party models and datasets, for example signed models?
Intake procedure; signature verification evidence
Insecure Supply Chains threat / B
AV-7.3
How do you vet third-party plugins, connectors and MCP servers offered in your marketplace?
Review process
Insecure Apps/Plugins threat / B
AV-7.4
What happens to our service if your upstream model provider fails, changes terms or is withdrawn?
Which logs can we access (prompts, outputs, tool calls, admin actions), in what format, and for how long? Can they stream to our SIEM?
Log schema; integration docs
Loss of Governance/Compliance threat / E
AV-8.2
Will log retention support our obligations, for example the EU AI Act Article 26(6) deployer duty to keep logs for at least six months for high-risk systems?
Retention settings
Loss of Governance/Compliance threat / E
AV-8.3
What is your contractual notification time for security incidents and AI-specific incidents affecting our data or actions taken on our systems?
Contract clause; incident policy
Service Failures threat / E
AV-8.4
Describe an AI-specific incident or vulnerability you handled in the last [24 months] and what changed as a result.
Control described but evidence missing, or covers only part of the scope
Request evidence or add a contract condition
0: Not met
Control absent or answer "unknown"
Finding: remediation plan, compensating control or risk acceptance
Blocker
Not met on AV-4.1, AV-5.2, AV-5.6 or AV-8.3 for a supplier handling confidential data or acting on our systems
Do not proceed until resolved or accepted by [CISO]
The vendor respondent signs: "The answers above are accurate for [service and version] as of [date]." The [third-party risk lead] records the overall rating, conditions and next review date.
Implementation guidance
Rollout
Pilot with three current AI suppliers of different types (SaaS assistant, model API, agent platform) and refine wording where answers were ambiguous.
Embed the questionnaire in the supplier onboarding workflow so that enabling a vendor's AI feature triggers it, not only new contracts.
Build a mapping sheet from these IDs to AI-CAIQ questions so vendors with a CSA submission answer once.
Feed blockers and conditions into contract templates: training-data exclusion, retention, notification time, log access, version-change notice.
Re-run on renewal, model change or new feature; keep prior answers to show drift.
Decision guidance
Self-attestation versus third-party evidence: a STAR for AI Level 1 entry is a self-assessment; an AIUC-1 certificate involves an accredited third-party audit and AIUC testing, and is valid for one year. Weigh evidence accordingly.
Small suppliers: accept shorter answers on sections 6 and 7 but never waive the blockers.
Large platforms: they rarely complete custom questionnaires; map their published documentation and certifications, then ask only the gaps.
Reference frameworks
Framework
What it contributes
CSA AI Controls Matrix (AICM) and AI-CAIQ
243 control objectives across 18 domains, a questionnaire for self-assessment or vendor evaluation, and mappings to ISO/IEC 42001, ISO 27001, NIST AI RMF, BSI AIC4 and the EU AI Act
AIUC-1
Agent-focused standard with six domains (A Data and Privacy, B Security, C Safety, D Reliability, E Accountability, F Society), independent audit and quarterly updates
ISO/IEC 42001:2023
AI management system certification for the supplier's governance
Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.
NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 27 October 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security.
nhimg.org
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.