Tokens & tokenization: the hidden attack surface

A 24-screen interactive deck: BPE, WordPiece and Unigram tokenizers trained live in your browser, then five families of tokenizer-level attacks (glitch tokens, homoglyphs, TokenBreak, token smuggling, token bombs) and nine hardening measures.

Reference Sennrich et al. (BPE) · Kudo (Unigram) · “Fishing for Magikarp” · TokenBreak (2025)