Tool Attack Defense on InjecAgent random architecture
7.3ASRPropGuard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PropGuardBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 7.3 | 96.7 | |
| Qwen3GuardBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 19.7 | 81.7 | |
| ThinkGuardBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 35.3 | 68.3 | |
| WildGuardBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 40.7 | 61.7 | |
| LlamaGuardBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 46.3 | 55 | |
| No DefenseBase Model=GPT-4o-mini, Communication Architecture=random2026.05 | 48.3 | 51.7 |