Prompt Injection Defense on GSM8K PI (test)
3.3ASR@1No Defense
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| No DefenseBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 3.3 | 7.7 | 11 | 96.7 | 93.3 | 90 | |
| AgentXposed-KickBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 3.3 | 8.7 | 10.7 | 96.7 | 91.7 | 90 | |
| INFA-GUARDBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 3.3 | 3.7 | 3.3 | 96.7 | 96.7 | 96.7 | |
| G-safeguardBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 3.7 | 3.3 | 3.3 | 96.7 | 96.7 | 96.7 | |
| AgentSafeBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 3.7 | 7.3 | 10.3 | 96.7 | 93.3 | 90 | |
| AgentXposed-GuideBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 4.3 | 5.7 | 7.3 | 95 | 95 | 93.3 | |
| ChallengerBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 5.5 | 14.5 | 15.9 | 93 | 84.7 | 84.7 | |
| No DefenseBackbone=GPT-4o-mini, Topology=Random2026.01 | 6.7 | 13 | 18 | 93.3 | 86.7 | 81.7 | |
| G-safeguardBackbone=GPT-4o-mini, Topology=Random2026.01 | 6.7 | 6.7 | 6.7 | 93.3 | 93.3 | 93.3 | |
| INFA-GUARDBackbone=GPT-4o-mini, Topology=Random2026.01 | 6.7 | 6.7 | 6.7 | 93.3 | 93.3 | 93.3 | |
| AgentSafeBackbone=GPT-4o-mini, Topology=Random2026.01 | 7 | 21.7 | 31.3 | 93.3 | 80 | 71.7 | |
| InspectorBackbone=Qwen3-235B-A22B, Topology=Random2026.01 | 7.2 | 10.3 | 16.1 | 92.3 | 92.3 | 82.1 | |
| ChallengerBackbone=GPT-4o-mini, Topology=Random2026.01 | 9 | 11.7 | 14.2 | 88.3 | 76.7 | 71.7 | |
| AgentXposed-GuideBackbone=GPT-4o-mini, Topology=Random2026.01 | 9.3 | 20 | 26.7 | 90 | 83.3 | 75 | |
| InspectorBackbone=GPT-4o-mini, Topology=Random2026.01 | 10.3 | 12.2 | 12.4 | 88.3 | 85 | 83.3 | |
| AgentXposed-KickBackbone=GPT-4o-mini, Topology=Random2026.01 | 10.7 | 26.3 | 35.7 | 88.3 | 73.3 | 65 |