A regex prefilter in front of an aligned model blocked 0% of attacks.

N=45 adversarial probes across three sub-corpora under adversarial mutation, and the primary substring classifier's block rate was 0% across all five OWASP LLM Top-10 categories tested (arXiv 2607.20494). The paper also reports metric-dependent divergence, meaning the apparent effectiveness of the regex layer changes with your measurement choice rather than with the actual defense. Small probe set, single source. But regex-then-model is genuinely everywhere in shipped LLM apps, and this says the first half is theater.