the fable 5 classifier: what benign prompt shapes are tripping it, post real repros

handed the ai security on top of my normal SRE load so im trying to actually characterize this thing instead of guessing. from what ive collected the false positives cluster on multi-step refactor asks and anything that reads like it touches auth or file deletion, even when its boring. not looking to dodge it, looking to document it so i can tell my devs what to avoid phrasing-wise. if you have a prompt that got flagged, post the shape (not the whole thing) and whether splitting it into smaller steps cleared it. assuming the classifier is dumber than advertised until someone shows me otherwise