Prompt Injection Defense on CSQA
13.4ASR@3INFA-GUARD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| INFA-GUARDBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 13.4 | 86.7 | |
| INFA-GUARDBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 23.3 | 76.7 | |
| G-safeguardBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 26.3 | 78.3 | |
| InspectorBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 26.9 | 73.1 | |
| G-safeguardBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 31.7 | 68.3 | |
| ChallengerBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 45.9 | 51.7 | |
| AgentXposed-KickBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 49 | 51.7 | |
| AgentXposed-GuideBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 55.3 | 46.7 | |
| AgentSafeBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 55.6 | 40 | |
| No DefenseBackbone=GPT-4o-mini, Communication Iterations=3, Topology Type=Random2026.01 | 59.3 | 41.7 | |
| ChallengerBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 72.8 | 18.3 | |
| AgentSafeBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 75.8 | 21.7 | |
| InspectorBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 76.3 | 21.7 | |
| AgentXposed-GuideBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 77 | 18.3 | |
| AgentXposed-KickBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 78 | 16.7 | |
| No DefenseBackbone=Qwen3-235B-A22B, Communication Iterations=3, Topology Type=Random2026.01 | 79.3 | 16.7 |