A Regex Filter in Front of an Aligned Model Blocked 0% Across All Five OWASP LLM Top-10 Categories Tested

arXiv 2607.20494·low signal

This study isolates the common production pattern of stacking a regex prefilter ahead of model-side alignment, evaluating against N=45 adversarial probes across three sub-corpora under adversarial mutation. Under the primary substring classifier the block rate is 0% across all five OWASP LLM Top-10 categories examined, and the paper reports metric-dependent divergence — meaning the apparent effectiveness of the regex layer changes with the measurement choice, not with the actual defense. It is a single-source finding with a small probe set, but it targets a defense pattern that is genuinely widespread in shipped LLM apps.