LexObf-12: Measuring and Hardening the Obfuscation Robustness of a Production Lexical Content Filter

Lexical filters remain the first stage of content moderation on many platforms because they are cheap, deterministic, and explainable, yet users evade them with substituted characters, inserted spaces, and look-alike Unicode letters. We introduce LexObf-12, a benchmark that applies twelve obfuscation classes to a filter's own lexicon in four carrier sentences, and use it, with two public labeled corpora and a list of common English words, to evaluate the pre-publication text filter of Kibhi (kibhi.com), a production short-video and messaging platform. The evaluation exposed two defects in the deployed normalizer: trailing punctuation was read as character substitution, so any listed phrase followed by "!" passed, and repeated-letter collapsing erased one listed term. Look-alike Cyrillic letters and zero-width characters defeated it entirely. After correction and hardening, mean detection across the twelve classes rose from 61.9% to 99.3%, against 17.7% and 28.7% for token and substring baselines, at 65 µs per check and with no change in false positives on 24,783 labeled tweets (1.95%) or among 10,000 common words. A leave-one-out ablation attributes the gain to each step and shows that only plural matching affects false positives. On 19,229 HateXplain posts, however, the filter flagged 12.0% of posts that annotators labeled normal, mostly through slur matches, and about 90% of the hate posts it missed contained no listed term at all. We conclude that normalization should be tested adversarially, that the limit of a word list is coverage rather than matching, and that slur matches should be held for human review rather than blocked. The benchmark generator is released.