Adversarial Attack on LLM Agents on GSM8K
97.71ASRInject Agent
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Inject AgentDefense Framework=No Defense2026.05 | 97.71 | — | 100 | 80 | |
| MINJADefense Framework=No Defense2026.05 | 68.86 | — | 74.29 | 60.57 | |
| MINJADefense Framework=Prompt Filter2026.05 | 65.43 | — | 72.86 | 57.71 | |
| OEPDefense Framework=No Defense2026.05 | 59.14 | 77.43 | — | 54.29 | |
| OEPDefense Framework=Prompt Filter2026.05 | 58.57 | 75.71 | — | 54.29 | |
| MemoryGraftDefense Framework=No Defense2026.05 | 52.57 | — | 100 | 47.71 | |
| MemoryGraftDefense Framework=Prompt Filter2026.05 | 50.57 | — | 100 | 48 | |
| AgentPoisonDefense Framework=Prompt Filter2026.05 | 48.86 | — | 100 | 41.14 | |
| AgentPoisonDefense Framework=No Defense2026.05 | 46.29 | — | 100 | 39.71 | |
| OEPDefense Framework=LLM Auditor2026.05 | 40.29 | 47.43 | — | 36.86 | |
| MINJADefense Framework=LLM Auditor2026.05 | 14.86 | — | 16.29 | 11.71 | |
| AgentPoisonDefense Framework=LLM Auditor2026.05 | 5.71 | — | 6.29 | 4.57 | |
| MemoryGraftDefense Framework=LLM Auditor2026.05 | 5.43 | — | 8.57 | 3.14 | |
| Inject AgentDefense Framework=LLM Auditor2026.05 | 3.71 | — | 4.86 | 1.43 | |
| Inject AgentDefense Framework=Prompt Filter2026.05 | 3.43 | — | 5.43 | 2.86 |