# 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 | 98.8% | 0.852 | | typo | 95.0% | 0.798 | | leet | 98.1% | 0.850 | | homoglyph | 99.4% | 0.878 | | zero_width | 100.0% | 0.885 | | case | 96.9% | 0.796 |