Agent Security (Indirect Prompt Injection) on AgentDojo Workspace (test)
2.1ASRBenchmark
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| BenchmarkModel=GLM-5.12026.05 | 2.1 | 75.7 | 0 | |
| AgentVigilModel=GLM-5.12026.05 | 2.1 | 94.3 | 0 | |
| IterInjectModel=GLM-5.12026.05 | 2.1 | 85.7 | 0 | |
| BenchmarkModel=MiniMax-M2.72026.05 | 5 | 87.9 | 0 | |
| IterInjectModel=MiniMax-M2.72026.05 | 7.1 | 92.9 | 1.4 | |
| AgentVigilModel=MiniMax-M2.72026.05 | 8.6 | 85 | 0 | |
| BenchmarkModel=Qwen3.5-27B2026.05 | 12.1 | 82.1 | 0.7 | |
| AgentVigilModel=Qwen3.5-27B2026.05 | 12.1 | 87.9 | 0.7 | |
| IterInjectModel=Qwen3.5-27B2026.05 | 13.6 | 85.7 | 0.7 | |
| BenchmarkModel=DS-V4-Flash2026.05 | 16.4 | 75 | 4.3 | |
| AgentVigilModel=DS-V4-Flash2026.05 | 19.3 | 73.6 | 0.7 | |
| IterInjectModel=DS-V4-Flash2026.05 | 21.4 | 84.3 | 6.4 |