Score a message for fraud, watch an attacker evade the detector, then watch a defense catch it back — live, in your browser.
loading model…
The trained tfidf-logreg baseline is strong on clean text. Swap to the keyword heuristic-v0 to see a brittle detector.
—
These character attacks run instantly in your browser. Set Defense → normalize to see whether input normalization catches the attack back: it reverses homoglyph and zero-width losslessly, but a semantic paraphrase would walk straight through it. LLM-based attacks live in the full API.