Security · Investigations · AIActive incident? Request help

Free working resource

AI security vendor review checklist

Turn broad assurances into a list of evidence to request. Use these twelve questions for a hosted service, local model or connected AI application. No email required.

This worksheet organizes your own review. It is not a security score, a technical test, a compliance assessment or a recommendation to approve a deployment. A “documented” answer records your assessment; it does not verify the evidence.

Answers are kept in your browser and are not submitted to this site. Use Reset answers to clear the worksheet. Export if you want to retain them. Ordinary site analytics are described in the privacy notice; worksheet selections are not sent to analytics.

Decision and ownership

Evidence to request: A named business task, intended users, prohibited uses and a decision owner.

Evidence to request: A record of what may enter prompts, uploads and retrieval, and who authorizes that use.

Evidence to request: The model, application, integrations and version covered by the review.

Data and access

Evidence to request: A map covering providers, logs, storage, backups, recipients and retention.

Evidence to request: Results showing what different user roles can access, including denied requests.

Evidence to request: The applicable terms and settings, with technical evidence where observable.

Connected actions

Evidence to request: Authorized tests using controlled content, recording expected and actual behavior.

Evidence to request: An inventory of tools, identities, privileges and enforcement outside the model.

Evidence to request: A tested approval boundary for sending, deleting, spending or changing access.

Evidence and follow-up

Evidence to request: A record of the setup, input, observed result, relevant version and limitations.

Evidence to request: Named contacts, escalation conditions, a stop mechanism and preserved logs.

Evidence to request: Owners, dates and a plan to recheck material changes to data, models or tools.

How to use the result

  1. Ask the system owner to identify the evidence behind each answer.
  2. Assign an owner and a review date to the material gaps.
  3. Agree which claims need authorized technical testing rather than another written assurance.
  4. Revisit the record when the model, tools, permissions or data use changes.

Sources and limits

These original review questions draw on the risk-management approach in the NIST AI RMF and the application-security issues described by OWASP. They are not an official NIST or OWASP assessment and do not imply endorsement. They address part of an AI review, not every risk.

Prepared by Joey Victorino. Published October 30, 2026. Share the public worksheet with colleagues; your selections are not included in the link.

Review the scope of an AI security engagement