Vendor AI Security Questionnaire
20 questions across 7 sections, adversarial in structure. Each question paired with an evidence requirement and a follow-up to use when the vendor declines to answer. Designed for procurement security review of third-party AI features and SaaS AI products — not a generic vendor risk questionnaire repurposed for AI.
Free spreadsheet
Excel · .xlsx · 1 sheet
Download free
Enter your work email. Send to the vendor; receive responses back; assess each response with the included evidence requirements and follow-up prompts. You'll also receive new blog posts when they publish.
Who it's for
Security architects, procurement leads, CISOs, and AI governance teams assessing a third-party AI feature or SaaS AI product before procurement or go-live.
Send it to the vendor, receive responses, and use the evidence requirements and follow-up prompts to assess each answer. The completed questionnaire feeds into the vendor assessment section of a Drel clearance review.
How to use it
- 1Send the questionnaire to the vendor with a response deadline and evidence-format expectations.
- 2Treat any refused or unanswered question as an evidence gap of equivalent weight to a 'no'.
- 3For each question, ask for the evidence — not just the answer.
- 4Use the follow-up prompts when the vendor answers ambiguously or declines.
- 5Treat any High-severity gap as a procurement blocker until a conditional clearance is agreed.
- 6Use the completed questionnaire as input to a vendor security review or AI Committee submission.
What's in the file
Seven sections, twenty questions total. Each question carries an evidence requirement and a follow-up prompt to use if the vendor answers “no” or declines.
| Section | Questions |
|---|---|
| Architecture | What models, where hosted, what tools and data sources accessed (3 questions) |
| Data | What data enters the model, training use, residency, retention (4 questions) |
| Model | Change management, OWASP LLM testing, model cards (3 questions) |
| Access | Internal access, SOC 2 scope, SSO support (3 questions) |
| Incident | Disclosure policy, incident history (2 questions) |
| Sub-processor | Sub-processor list, contractual flow-down (2 questions) |
| Re-assessment | Change notifications, security review cadence, right to audit (3 questions) |
From questionnaire to review pack
Spreadsheets are the starting point.
This questionnaire structures the vendor evidence you need before a go-live decision. Drel turns that evidence into a guided AI security review — mapping vendor responses to control gaps, generating a risk disposition, and producing a review-ready dossier your AI Committee can actually approve or reject.
Frequently asked
Frequently asked questions
- Why an adversarial questionnaire?
- Extractive questionnaires ask the vendor to describe their AI feature. Vendors return marketing copy. Adversarial questionnaires ask for the description and then ask for the evidence that proves it — and they include a follow-up to use when the vendor declines or answers ambiguously. The output is a structured record of what the vendor disclosed and what they refused to disclose.
- What if the vendor refuses to answer some questions?
- A refusal is itself an evidence gap and goes into the assessment record at equivalent weight to a 'no'. The conditional clearance you produce can require the vendor to provide the missing answers before moving to production. A vendor that consistently refuses produces a control gap the AI Committee must explicitly accept or reject.
- How is this different from a generic SIG/CAIQ questionnaire?
- SIG and CAIQ cover information security generally and do not address AI-specific concerns — training data use, model change management, prompt/output handling, AI-specific incident definitions. This questionnaire is AI-specific and pairs each question with both an evidence requirement and an AI-specific follow-up.
- Should this replace our SIG/CAIQ?
- No. It complements them. Use SIG/CAIQ for general information security; use this for AI-specific concerns. Together they cover the full vendor security posture.
- Can we send this directly to the vendor?
- Yes. The spreadsheet is designed for that workflow: you send it, the vendor fills in the response columns, and you assess each response with the internal notes column.
- How does this map to EU AI Act deployer obligations?
- Under the EU AI Act, deployers of high-risk AI systems have evidence obligations even for vendor-supplied AI. The questionnaire produces the procurement-time evidence that supports those obligations: documented assessment of the vendor system, identified gaps, and conditions for use.