Prompt Injection Detection on VPI-Bench
87.58RecallWebAgentGuard-8B
Evaluation Results
| Method | Links | |
|---|---|---|
| WebAgentGuard-8BModel Type=Ours2026.04 | 87.58 | |
| WebAgentGuard-4BModel Type=Ours2026.04 | 85.95 | |
| GPT-4oModel Type=Closed-source APIs2026.04 | 78.43 | |
| GPT-4o-MiniModel Type=Closed-source APIs2026.04 | 69.93 | |
| GPT-4.1Model Type=Closed-source APIs2026.04 | 69.61 | |
| Llama-Guard-3-Vision-11BModel Type=Guard models2026.04 | 68.95 | |
| Prompt-Guard-1-86MModel Type=Guard models2026.04 | 64.38 | |
| Llama-3.2-Vision-Instruct-11BModel Type=Open-source instructed models2026.04 | 43.79 | |
| Qwen3-VL-Instruct-8BModel Type=Open-source instructed models2026.04 | 42.16 | |
| Qwen3-VL-Instruct-4BModel Type=Open-source instructed models2026.04 | 40.52 | |
| Qwen2.5-VL-Instruct-7BModel Type=Open-source instructed models2026.04 | 14.71 | |
| GuardReasoner-VL-7BModel Type=Guard models2026.04 | 4.9 | |
| Prompt-Guard-2-86MModel Type=Guard models2026.04 | 0 |