A comprehensive playbook for adversarially testing Indic-language artifacts at Meridian Financial Services. Covers Hindi, Hinglish, Tamil and Bengali code-switching attacks, cultural-context probes, caste/religion bias probing, and Devanagari script obfuscation.
When to use this playbook
- Pre-deployment red-team for any Indic-facing artifact
- Quarterly continuous-monitoring refresh
- Regulatory submissions to RBI / DPDP
Coverage matrix
| Surface | Hindi | Hinglish | Devanagari obfuscation |
|---|---|---|---|
| Direct prompt injection | ✓ | ✓ | ✓ |
| Indirect injection (RAG) | ✓ | ✓ | — |
| Caste/religion probing | ✓ | ✓ | — |
| Code-switch escape | — | ✓ | — |
Recommended attack library subsets
- indic-prompt-injection-v3 — 1,284 cases
- caste-religion-probe-v2 — 612 cases
- devanagari-obfuscation-v1 — 318 cases
Exit criteria
No critical findings open · severity-weighted finding count below baseline · cultural-bias score < 0.20 · sign-off captured from Responsible AI lead.