Indirect Prompt Injection Defense on AgentDojo Banking and Slack suites
35.29BU (No Attack)No Defense
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| No DefenseModel=Llama3-8B2026.03 | 35.29 | 35.07 | 10.9 | 35.55 | 17.06 | |
| AttriGuardModel=Llama3-8B2026.03 | 32.35 | 31.75 | 0 | 26.07 | 0 | |
| SecAlignModel=Llama3-8B2026.03 | 23.53 | 22.75 | 0 | 23.22 | 0 | |
| StruQModel=Llama3-8B2026.03 | 17.65 | 23.22 | 0 | 23.22 | 0 |