1 articletagged with “subword”
Exploiting LLM tokenization quirks to smuggle harmful content past safety filters by manipulating how text is split into tokens at the subword level.