Prompt Injection on MATH
6Attack Success Rate (ASR)PropGuard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PropGuardArchitecture=Random2026.05 | 6 | 95 | |
| PropGuardArchitecture=Tree2026.05 | 7 | 93.3 | |
| PropGuardArchitecture=Chain2026.05 | 8 | 91.7 | |
| PropGuardArchitecture=Star2026.05 | 11 | 90 | |
| G-SafeguardArchitecture=Tree2026.05 | 19 | 81.7 | |
| GUARDIANArchitecture=Tree2026.05 | 19.3 | 83.3 | |
| GUARDIANArchitecture=Chain2026.05 | 20 | 81.7 | |
| GUARDIANArchitecture=Star2026.05 | 20.3 | 80 | |
| GUARDIANArchitecture=Random2026.05 | 20.3 | 81.7 | |
| INFA-GuardArchitecture=Random2026.05 | 21.7 | 80 | |
| G-SafeguardArchitecture=Chain2026.05 | 23 | 78.3 | |
| No DefenseArchitecture=Tree2026.05 | 23.3 | 76.7 | |
| G-SafeguardArchitecture=Star2026.05 | 23.3 | 75 | |
| INFA-GuardArchitecture=Chain2026.05 | 24.3 | 76.7 | |
| INFA-GuardArchitecture=Star2026.05 | 24.3 | 76.7 | |
| No DefenseArchitecture=Chain2026.05 | 24.7 | 75 | |
| AgentSafeArchitecture=Tree2026.05 | 25 | 75 | |
| AgentSafeArchitecture=Star2026.05 | 25 | 75 | |
| AgentSafeArchitecture=Chain2026.05 | 26 | 76.7 | |
| No DefenseArchitecture=Star2026.05 | 26 | 75 | |
| G-SafeguardArchitecture=Random2026.05 | 26 | 75 | |
| INFA-GuardArchitecture=Tree2026.05 | 26.7 | 75 | |
| No DefenseArchitecture=Random2026.05 | 26.7 | 71.7 | |
| AgentSafeArchitecture=Random2026.05 | 32.7 | 66.7 | |
| AgentXposedArchitecture=Random2026.05 | 36.3 | 65 | |
| ChallengerArchitecture=Random2026.05 | 37 | 63.3 | |
| AgentXposedArchitecture=Chain2026.05 | 37.7 | 63.3 | |
| InspectorArchitecture=Star2026.05 | 38.7 | 60 | |
| InspectorArchitecture=Chain2026.05 | 39.7 | 61.7 | |
| ChallengerArchitecture=Star2026.05 | 40 | 60 | |
| ChallengerArchitecture=Tree2026.05 | 40.7 | 58.3 | |
| ChallengerArchitecture=Chain2026.05 | 41 | 58.3 | |
| AgentXposedArchitecture=Star2026.05 | 41.3 | 60 | |
| InspectorArchitecture=Random2026.05 | 41.3 | 60 | |
| InspectorArchitecture=Tree2026.05 | 42.3 | 58.3 | |
| AgentXposedArchitecture=Tree2026.05 | 43 | 56.7 |