# Adversarial robustness (unsafe items under text attacks) n=160 unsafe items, threshold 0.5, detection rate = fraction still flagged unsafe. | Attack | Detection rate | Mean unsafe prob | | --- | --- | --- | | clean | 100.0% | 0.890 | | typo | 92.5% | 0.826 | | leet | 96.9% | 0.869 | | homoglyph | 100.0% | 0.922 | | zero_width | 100.0% | 0.914 | | case | 96.2% | 0.817 |