Every prompt, inspected before it reaches the model.
Drop LLMFirewall between your app and Gemini. It scans every outbound prompt for SSNs, card and IBAN numbers, emails, and injection attempts — including hidden Unicode used to smuggle content past a naive filter — blocks what it finds, re-checks the model's reply, and caps spend so one runaway loop can't drain your budget.
- SSNs
- Card & IBAN numbers
- Emails
- Prompt injection
- Hidden Unicode
- Budget overrun
Every request, checked twice
Pre-check the prompt
Regex detection for SSN-like patterns, card numbers (Luhn-validated), IBANs (mod-97 validated), emails, phone numbers, known injection phrases tolerant of spacing tricks, and hidden Unicode used to smuggle content past a naive filter. A match blocks the request — Gemini is never called. No model makes this call, so there's nothing here for an adaptive attacker to argue with — see exactly what runs →
Call Gemini
Clean prompts go through to gemini-3.1-flash-lite. A per-IP hourly cap and a global
daily cap sit in front of every call, so one bad actor can't blow the whole budget.
Post-check the reply
The model's own response is scanned with the same PII detectors before it's returned — catching the case where Gemini repeats something sensitive back.
What the firewall has caught recently
The last 20 checks from every visitor to this page, rolling 24-hour window — finding type and time only, never the prompt itself. Quiet stretch here just means nobody's run a check in the last day, not that anything broke.
| Time | Verdict | Finding types |
|---|
No activity yet — be the first, run a prompt above.
Want this in front of your own Gemini traffic?
This build protects Gemini only and doesn't issue API keys yet — leave your email and we'll reach out when self-serve access opens.