Agent Security (Indirect Prompt Injection) on AgentDojo Banking (test)
63.8ASRIterInject
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| IterInjectModel=Qwen3.5-27B2026.05 | 63.8 | 46.9 | 16.2 | |
| AgentVigilModel=Qwen3.5-27B2026.05 | 63.1 | 49.2 | 19.2 | |
| IterInjectModel=DS-V4-Flash2026.05 | 60.8 | 68.5 | 33.8 | |
| AgentVigilModel=DS-V4-Flash2026.05 | 57.7 | 32.3 | 13.8 | |
| BenchmarkModel=Qwen3.5-27B2026.05 | 52.3 | 46.9 | 16.2 | |
| IterInjectModel=MiniMax-M2.72026.05 | 47.7 | 29.2 | 6.9 | |
| BenchmarkModel=DS-V4-Flash2026.05 | 47.7 | 66.9 | 30.8 | |
| AgentVigilModel=GLM-5.12026.05 | 44.6 | 48.5 | 14.6 | |
| AgentVigilModel=MiniMax-M2.72026.05 | 44.6 | 26.2 | 5.4 | |
| IterInjectModel=GLM-5.12026.05 | 42.3 | 46.9 | 16.2 | |
| BenchmarkModel=GLM-5.12026.05 | 24.6 | 40 | 10.8 | |
| BenchmarkModel=MiniMax-M2.72026.05 | 24.6 | 27.7 | 5.4 |