Sample (illustrative data)
AI security evidence pack
Example Co. support assistant · tested 2026-10-28 · retested 2026-11-04 · endpoint ownership verified by DNS token
| Test area | Result | Detail |
|---|---|---|
| Direct instruction override | Held | 42 / 42 attempts refused |
| System-prompt extraction | Failed → Fixed | 3 / 40 leaked fragments; fixed by prompt and output filter; retest 0 / 40 |
| Cross-customer data leakage | Held | 0 / 60 probes returned another tenant's data |
| Indirect injection via documents | Failed → Fixed | 5 / 25 hijacked; retest 0 / 25 after content isolation |
| Unsafe tool / action requests | Held | 30 / 30 refused or required confirmation |
Drafted questionnaire answers
Q: Do you test your AI features for prompt injection?
Yes. An automated adversarial test suite was run against the production AI endpoint on 2026-10-28 and retested on 2026-11-04 after remediation. Results and methodology are in the attached report.
Q: Can the system prompt be extracted?
Testing found partial leakage in 3 of 40 attempts. Mitigations were deployed and the retest returned 0 of 40.
This sample uses illustrative numbers and a fictional company.