This does surprise me, because you'd think that even if they crank up the filter's sensitivity at the expense of specificity, an LLM company wouldn't simply design a filter that triggers on keywords in a completely unrelated context.