Template Library
AI Security Controls Checklist
An Excel workbook for AI security readiness: twelve controls scored 0 to 2 with gates, evidence, owners, and dates, a use check that flags private data plus untrusted content plus outside action, and a dashboard with a result.
The Controls sheet lists twelve controls: approved tools and shadow AI, data classes mapped to tools, sign-in and access control, least privilege for agents, secrets kept out of prompts, prompt-injection controls, logging and retention, vendor security review, abuse-case testing, incident response, separation from control systems, and training. Score each 0, 1, or 2, and record the evidence, an owner, a next step, a due date, and when it was last reviewed. Three controls are gates, and a flag shows gate failures, overdue actions, and items that need a plan.
The Use Check sheet lists each AI use with its data class and whether it reads private data, reads untrusted content, and can send or act outside. It counts the legs of the pattern, flags uses with all three, and also tracks agent identity, whether a person approves outside actions, logging, and security review dates. The Dashboard gives the total score, the result (stop, weak, gaps to close, or strong), and counts of uses that need attention.
Example scores and six example uses at an illustrative plant are included so you can see the formulas working before you replace them. This is a planning and awareness aid, not a security assessment.
What Is Included in the Workbook
| Sheet | Purpose | What Teams Capture |
|---|---|---|
| Start Here | Orientation and workflow | What the workbook does and a numbered workflow with a link to the sheet for each step |
| Dashboard | Live summary of security readiness | Total score, controls scored, gate items at zero, the result, uses with all three legs, shared logins, controls by score and flag, uses by measure, and a chart of control scores |
| Controls | The twelve controls (Excel Table) | Control, what good looks like, gate, score, evidence, owner, next step, due date, last reviewed, and a calculated flag |
| Use Check | AI uses and the risk pattern (Excel Table) | Use, owner, tool or vendor, data class, reads private data, reads untrusted content, can act outside, legs of the pattern, agent identity, approval before outside action, logged, last security review, and a calculated review check |
| Settings | Thresholds, as-of date, and lists | The totals for strong and for gaps to close, the as-of date, the review interval, and the lists for data class, yes or no, agent identity, and yes, no, or not applicable |
Key Features Inside the Template
Gates That Stop a Go
A zero on approved tools, data classes, or least privilege stops the decision.
Evidence, Owners, and Dates
Every score carries evidence and a next step.
A Check on Each Use
Flags uses that read private data and outside content and can act outside.
A Result You Can Show
Strong, gaps to close, weak, or stop, from a rule you can edit.
Best Use Cases
- Checking the security basics before launching an AI use or agent
- Reviewing the AI estate once or twice a year
- Preparing for a conversation with IT and security leads
- Showing a governance group or auditor what is in place
- Teaching AI security awareness to operations leaders
How to Use the Template Effectively
- Check the thresholds and the as-of date on the Settings sheet.
- Score each control from evidence, with your IT and security leads. Add the evidence, owner, next step, and date.
- List each AI use on the Use Check sheet and answer the three questions about it.
- Close the gate gaps first, then the lowest scores. Remove a leg from any use that has all three, or put a person in front of the outside action.
- Have qualified security specialists review important uses and test them.
- Score again after any significant change, and at least once a year.
Who Should Use This Template
- Operations, quality, and safety leaders
- IT and security teams
- AI governance groups
- Students and trainees learning AI risk
Common Mistakes to Avoid
- Scoring from hope instead of evidence
- Treating a high score as proof of security
- Relying on one layer, such as a filter
- Agents on shared accounts with broad access
- Never testing the kill switch or the abuse cases
Important Notes
AI Security Controls Checklist Frequently Asked Questions
Does a high score mean we are secure?
No. The scores are your own judgments and nothing is tested. Have qualified security specialists review important uses and test them.
Why are three controls gates?
Approved tools, data classes mapped to tools, and least privilege for agents limit what can leak and what a mistake or attack can do. If they are not in place, the other controls work against a larger exposure.
What is the risk pattern on the Use Check sheet?
A use that reads private data, reads untrusted content, and can send or act outside is the most exposed, because an attacker only has to get instructions into the content. Remove one of the three or put a person in front of the action.
Is this a substitute for a security assessment?
No. It is a planning and awareness aid. Security specialists should assess and test important uses.