// tok3n_inflat0r.exe

Adversarial Tokenmaxxing

LLMs rely on BPE tokenizers to convert spans of raw text into their input tokens. These tokenizers merge common character chains into single tokens. This demo finds same-length lookalike substitutions (digits, Cyrillic homoglyphs, fullwidth forms, case flips) that break those merges, maximizing token count while keeping character length fixed. Same length readable(ish) text, but many more tokens, resulting in a much higher cost for any LLM to read it.