Prompt Injection Defense Evaluation on AgentDojo without attack
0.49Total Tokens (M)tool_filter
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| tool_filterDefense Method=tool_filter2025.06 | 0.49 | 50.4 | 7.6 | 86.6 | |
| undefended agentDefense Method=undefended agent2025.06 | 0.82 | 48.3 | 30.7 | 21.4 | |
| spotlighting_with_delimitingDefense Method=spotlighting_with_delimiting2025.06 | 0.88 | 41 | 41.8 | -0.9 | |
| DRIFTDefense Method=DRIFT2025.06 | 2.37 | 50.9 | 1.4 | 20.9 | |
| transformers_pi_detectorDefense Method=transformers_pi_detector2025.06 | 2.58 | 21.2 | 13 | 3.2 | |
| ProgentDefense Method=Progent2025.06 | 2.6 | 45.6 | 9.4 | 13.9 | |
| repeat_user_promptDefense Method=repeat_user_prompt2025.06 | 5.43 | 47.1 | 15.5 | 5.8 | |
| CaMeLDefense Method=CaMeL2025.06 | 6.09 | 35.4 | 0 | 5.8 |