Prompt Injection Defense on AgentDojo Travel suite v1 (test)
20CUTool Filter
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Tool FilterTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 20 | 18.57 | 0.71 | |
| Repeat PromptTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 20 | 12.86 | 0 | |
| DelimitingTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 20 | 17.14 | 5.71 | |
| PI DetectorTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 20 | 7.14 | 0 | |
| Task ShieldTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 20 | 10.71 | 0 | |
| No DefenseTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 15 | 17.86 | 1.43 |