Prompt Injection Defense on AgentDojo Banking suite v1 (test)
62.5CUTask Shield
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Task ShieldTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 62.5 | 43.75 | 4.17 | |
| PI DetectorTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 43.75 | 36.11 | 8.33 | |
| No DefenseTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 37.5 | 32.64 | 25.69 | |
| Tool FilterTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 37.5 | 36.11 | 4.17 | |
| Repeat PromptTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 37.5 | 31.25 | 12.5 | |
| DelimitingTarget Model=GPT-3.5-turbo, Attack=Important Messages2024.12 | 37.5 | 34.72 | 25.69 |