Your AI filter is fluent in English. Attackers aren't.
Standard injection detectors are trained on clean English text. Real attacks use leetspeak and casual gaul Indonesian — and slip right through. Paste a prompt and watch it happen, live.
Try a preset (★ = where English detectors go blind):
Both are real, live models. Nothing you type is stored.
What you're seeing
Two real detectors, same prompt
The left column is ProtectAI's DeBERTa v3 — the popular open-source English injection detector. The right is Anoman, tuned for English, Bahasa Indonesia, and obfuscation. Try the ★ presets: a jailbreak written in leetspeak (4l4y) or casual gaul slips past the English-only model, while Anoman flags it. The plain-attack presets are caught by both — Anoman covers those too.
Catch the attacks English-only filters miss
Prompt-injection detection tuned for how attacks actually look in Indonesia — on every call, billed in Rupiah, processed in Jakarta.