Agentic Prompt Injection Defense on AgentDojo (test)
0ASR (Direct)MELON
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MELONDefensive Paradigm=SOTA runtime verification2025.12 | 0 | 0 | 0 | 0.63 | 0.16 | 42.36 | 49.42 | 47.84 | 35.41 | 43.76 | 72.16 | |
| CCAFull Name=Cognitive Control Architecture2025.12 | 0.42 | 0 | 0.11 | 0.84 | 0.34 | 85.35 | 87.88 | 85.56 | 86.93 | 86.43 | 87.63 | |
| DeBERTaDefensive Paradigm=External model-based classifier2025.12 | 0.74 | 0.11 | 0.63 | 7.59 | 2.268 | 51.84 | 35.62 | 45.63 | 38.78 | 42.97 | 53.61 | |
| Repeat PromptDefensive Paradigm=Prompt engineering-based2025.12 | 0.95 | 0.21 | 0.32 | 18.44 | 4.98 | 67.54 | 65.23 | 65.02 | 54.69 | 63.14 | 59.79 | |
| No Defense2025.12 | 2.32 | 1.69 | 2.63 | 41.31 | 11.99 | 87.46 | 86.09 | 86.51 | 66.6 | 81.67 | 89.69 | |
| SpotlightDefensive Paradigm=Prompt engineering-based2025.12 | 3.16 | 1.16 | 2.32 | 42.78 | 12.36 | 88.09 | 88.09 | 87.67 | 66.81 | 82.67 | 91.75 |