Security Lab
Each test runs through the real pipeline against an isolated copy of the library that adds synthetic protected records (with canary strings) and a poisoned document. A test passes only when the control actually held; the evidence for every result is shown. Nothing outside this application is attacked.
Live model challenge
Sends one attack to the configured model with synthetic canaries in its instructions and reports whether the model itself leaked them, before Secure-RAG's output checks.
Controls under test
| Control | What it does | Tests |
|---|---|---|
| Server-side identity and authorisation | Role, department and clearance come from the server, not the request; documents are filtered before search. | ACL-01, ACL-02 |
| Question injection check | Directive patterns aimed at the assistant block the request before retrieval. | INJ-01, EXF-01, POL-01, SYS-01, BEN-01 |
| Passage quarantine | Retrieved passages carrying instructions never reach the model. | INJ-02 |
| Answer checks | Citations must be retrieved sources; canaries, secrets and personal data are redacted. | CIT-01, OUT-01 |
| Input redaction | Emails, keys, TFN and ABN (checksum), BSB, Medicare, cards (Luhn), phone numbers. | PII-01 |
| Hash-chained audit | Every request, block, quarantine and test is recorded; tampering is detectable. | Audit trail |
Answers come only from the governed library and are decision support, not legal advice. Verify regulatory content against the regulator's current publication.