Prompt Injection Detection on AgentDojo
97Banking RLRedVisorLlama
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| RedVisorLlamaBackbone=Llama-3-8B, Attack Type=Important Instructions2026.02 | 97 | 97 | 90 | 90 | 88 | 89 | 89 | 91 | |
| PromptLocateAttack Type=Important Instructions2026.02 | 96 | 97 | 87 | 89 | 81 | 84 | 86 | 91 | |
| RedVisorQwenBackbone=Qwen-2.5-7B, Attack Type=Important Instructions2026.02 | 95 | 96 | 86 | 89 | 80 | 81 | 84 | 87 | |
| RedVisorMistBackbone=Mistral-7B, Attack Type=Important Instructions2026.02 | 94 | 95 | 83 | 85 | 78 | 80 | 82 | 85 | |
| PromptArmorAttack Type=Important Instructions2026.02 | 11 | 15 | 8 | 10 | 6 | 7 | 9 | 11 |