Skill Invocation Safety Auditing on SIA-Bench Held-out indirect prompt injection
43.9HR-AUPRCFusion
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Fusion2026.04 | 43.9 | 15.2 | 55.6 | 0.172 | 29.8 | 0.395 | |
| Contextual2026.04 | 40.5 | 14.6 | 53.3 | 0.127 | 28.9 | 0.38 | |
| Text-Only2026.04 | 39.9 | 11 | 40 | 0.125 | 30.4 | 0.38 | |
| Agent-Audit2026.04 | 38 | 11.4 | 41.6 | 0.164 | 43.5 | 0.435 | |
| No Audit2026.04 | 36.4 | 10 | 36.4 | 0 | 43.8 | 0.438 | |
| Static Prior2026.04 | 33.9 | 2.3 | 8.3 | 0.011 | 29.8 | 0.391 | |
| Denylist2026.04 | 33.9 | 2.3 | 8.3 | 0.011 | 29.8 | 0.391 |