# Employee AI security evaluation checklist

Version: 19 September 2026. Synthetic tests only. This is an evaluation worksheet, not a completed test or coverage certification.

## Before the test

- Use an approved non-customer environment and a dedicated test group.
- Record the AI product and interface, OS, browser, installed Aona version and policy version.
- Record backend hosting and prompt-processing placement separately.
- Agree the action to test: block, redact or another explicitly supported policy response.
- Use the supplied fictional examples, never customer prompts, secrets or documents.
- Define pass, fail and not-tested before running the test. An empty result means not tested.

## Prompt checks

1. Submit a clean synthetic prompt as a negative control.
2. Test typed and pasted sensitive prompts separately.
3. For block: confirm the request is not submitted and the employee has no continue override.
4. For redact: compare original and submitted content, the employee notice and the retained useful task.
5. Confirm the event identifier and policy outcome for that same attempt.
6. Repeat on each app and client in scope. A result on ChatGPT does not establish Claude or desktop coverage.

## Document checks

1. Build the sample DOCX from file-test-recipe.md. Keep an unchanged original.
2. Test attachment button, drag-and-drop and multiple-file paths separately where supported.
3. Compare the original, returned document and submitted content.
4. Check names, email addresses, account fields, paragraphs and table layout.
5. Test XLSX formulas, hidden sheets/comments and PDF text/scanned pages separately. Do not assume support from the extension alone.
6. Record rejection, failure and recovery behavior for unsupported or malformed content.

## Sentinel checks

1. Record the connector artifact/version, API contract and test workspace.
2. Trace one synthetic EventOriginalUid from the source API to AonaEvents_CL and Aona().
3. Compare EventTime, Disposition and the query result. TargetAppName is Aona in the current mapping.
4. Check overlapping-poll deduplication, missing periods, pagination and event completeness.
5. A local fixture or parser check is not a live Sentinel-ingestion result.

## Decision record

Use test-record.json. Capture evidence location, observed result, exclusions, reviewer and next action. Record not-tested explicitly. Remove customer identifiers and credentials from every artifact before sharing it.

Read current coverage: https://aona.ai/resources/ai-security-coverage/
Read deployment options: https://aona.ai/resources/ai-security-deployment/
Discuss your security evaluation: https://aona.ai/book-demo/