Know where your Persian AI agent breaks.
Focused, source-linked tests for Persian chatbots, RAG systems and tool-using agents. Get reproducible cases, prioritized findings and a written retest—not an empty safety score.
A compact, actionable report—not a certification.
10 agreed test scenarios
Results + reproducible prompts
Prioritized findings + one written retest
48 real benchmark cases. Public scripts. Traceable results.
PAIB is a Romo7 research/engineering project, not a customer engagement. The suite tests six areas: instruction hierarchy, indirect injection, simulated secret protection, tool authorization, evidence integrity and cross-lingual attacks. All public tool actions and secrets are fictional.
Exploratory run, 1 Oct 2026
| Model | Coverage | Security outcome | Format |
|---|---|---|---|
| Poolside Laguna S 2.1 | 97.9% | 70.2% | 59.6% |
| NVIDIA Nemotron 3 Ultra 550B | 100% | 66.7% | 56.3% |
The results describe only the public suite and tested configurations; they do not rank models universally or prove a deployed agent is secure.
How an engagement works
- Agree on one use case and allowed test boundary
- Run only authorized, non-destructive scenarios
- Deliver evidence, severity and corrective recommendations
- Retest after the fix and document remaining limits
Other work samples
Start a conversation
Tell us what the bot does, the main failure you worry about, and whether you have a safe demo. No secrets or production credentials. We will confirm scope, eligibility and lawful payment before any work.
The GitHub issue form is public and requires GitHub sign-in. Do not share credentials, customer records or confidential prompts. You can start privately via X instead. There is no payment gateway or hidden auto-submission on this page.