AI Security Checklist: a 30-point self-audit
Thirty concrete controls across inputs, tools, data, outputs, operations and governance. Tick what you have, see your score and the gaps that matter most. Nothing leaves your browser.
- 30 checks, 6 categories
- About 10 minutes
- Saves locally in your browser
Tick the controls you have in place to see your score.
Inputs and prompts 0/5
Tools and agency 0/5
Data and retrieval 0/5
Outputs 0/5
Monitoring and operations 0/5
Governance and supply chain 0/5
What to fix first
Open high-impact items, in order. Each is a design change you can usually make without a big project.
- All 14 high-impact items are open. Start with tools and agency.
Scoring: high-impact items count 2 points, others 1, out of 44. Your ticks are saved in this browser only.
What the bands mean
| Score | Band | What it usually means | Suggested next step |
|---|---|---|---|
| 0–39 | Exposed | Core boundaries are missing: a successful injection or a looping agent would cause real damage. | Fix the open high-impact items in tools, outputs and operations this sprint. Consider an audit. |
| 40–69 | Partial | Some layers exist but gaps remain, often in retrieval permissions, logging or governance. | Close the high-impact list, then add regression probes to CI. |
| 70–89 | Maturing | Defence in depth is mostly in place. Remaining items are hygiene and evidence. | Validate with an AI red team; formalise governance. |
| 90–100 | Strong | Controls are in place across all six categories. | Prove they work: red team annually and keep evals and probes running on every change. |
Where the list comes from
The items are the controls we check first in every LLM security audit, grouped by where they sit in the request path. They map to the OWASP Top 10 for LLM Applications without quoting it, because the question that matters is not "which category?" but "what did you build?"
High-impact items are the ones that, when missing, turn a nuisance into an incident: an injection becomes a data breach because the tool ran as a service account, or a bug becomes a large bill because there was no spend cap. The rest reduce likelihood or improve your evidence.
For the reasoning behind each item, read the long-form LLM security checklist guide, the prompt injection defence page, or the glossary entry. If you build agents, the AI agents and automation page shows how we design the tools layer from the start.
Frequently asked questions
Is my checklist data sent to Teksmobile?
No. Your ticks are stored only in your browser's local storage so they survive a page reload. Clearing your browser data or pressing "Clear all" removes them. Nothing is transmitted.
How is the score calculated?
High-impact items count two points, others one, for a total of 44 points. The score is the percentage of points you have ticked, rounded. The bands are: under 40 exposed, 40 to 69 partial, 70 to 89 maturing, 90 and above strong.
What should I do with a low score?
Work through the "what to fix first" list, which shows your open high-impact items in order. Most are design changes rather than big projects. If you want the findings verified against the real system, ask for an LLM security audit.
Does a perfect score mean we are secure?
It means you have the controls we consider essential. It does not prove they work. An AI red team is how you test that.