Adversarial Robustness on Surface-form evasion probe suite Paraphrase-sensitive spans
47.6ExposureRegex-only extractor
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Regex-only extractorType=Baseline2026.04 | 47.6 | — | |
| BodhiPromptShieldPolicy Mediator=Normalization-aware2026.04 | 47.6 | 50 |