Prompt Injection and Jailbreak Detection on WARD (test)
99.33AccuracyWARD-0.8b
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| WARD-0.8bModel Category=Ours, Number of Parameters=0.8b2026.05 | 99.33 | 99.6 | 99.07 | 99.34 | |
| WARD-2bModel Category=Ours, Number of Parameters=2b2026.05 | 99.2 | 99.67 | 98.75 | 99.2 | |
| Gemini-3-FlashModel Category=Closed-source API2026.05 | 96.37 | 96.53 | 96.21 | 96.37 | |
| Gemma-4-31BModel Category=Open-source instructed model, Number of Parameters=31B2026.05 | 95.5 | 91.67 | 99.28 | 95.32 | |
| GPT-5.4-MiniModel Category=Closed-source API2026.05 | 95.27 | 96.6 | 94.09 | 95.33 | |
| Gemini-3.1-proModel Category=Closed-source API2026.05 | 94.7 | 96.2 | 93.4 | 94.78 | |
| GPT-5.4Model Category=Closed-source API2026.05 | 93.47 | 92.07 | 94.72 | 93.37 | |
| WebAgentGuard-8bModel Category=Guard model (Prompt Injection), Number of Parameters=8b2026.05 | 91.97 | 85.2 | 98.54 | 91.38 | |
| Claude-Haiku-4.5Model Category=Closed-source API2026.05 | 90.86 | 98.55 | 84.49 | 90.98 | |
| Claude-Sonnet-4.6Model Category=Closed-source API2026.05 | 90.43 | 97.4 | 85.49 | 91.06 | |
| WebAgentGuard-4bModel Category=Guard model (Prompt Injection), Number of Parameters=4b2026.05 | 88.3 | 78.4 | 97.76 | 87.01 | |
| Qwen-3.5-4BModel Category=Open-source instructed model, Number of Parameters=4B2026.05 | 87.37 | 86.47 | 88.05 | 87.25 | |
| Qwen3.5-35B-A3BModel Category=Open-source instructed model, Number of Parameters=35B2026.05 | 86.03 | 85.53 | 86.4 | 85.96 | |
| Qwen-3.5-2BModel Category=Open-source instructed model, Number of Parameters=2B2026.05 | 82.5 | 85.2 | 80.83 | 82.96 | |
| Gemma-3-4BModel Category=Open-source instructed model, Number of Parameters=4B2026.05 | 78.67 | 84.67 | 75.6 | 79.87 | |
| GPT-Oss-Safeguard-20bModel Category=Guard model (Jail Break), Number of Parameters=20b2026.05 | 70.33 | 40.67 | 100 | 57.82 | |
| Qwen-3.5-0.8BModel Category=Open-source instructed model, Number of Parameters=0.8B2026.05 | 68.4 | 73.67 | 66.65 | 69.98 | |
| SuperAgent-Guard-4bModel Category=Guard model (Prompt Injection), Number of Parameters=4b2026.05 | 68.4 | 37.07 | 99.29 | 53.98 | |
| Llama-Guard-4-12BModel Category=Guard model (Jail Break), Number of Parameters=12B2026.05 | 61.27 | 23.2 | 97.21 | 37.46 | |
| BrowseSafeModel Category=Guard model (Prompt Injection)2026.05 | 59.7 | 19.4 | 100 | 32.5 | |
| Prompt-Guard-1-86MModel Category=Guard model (Jail Break), Number of Parameters=86M2026.05 | 57.43 | 32.13 | 65.05 | 43.02 | |
| SuperAgent-Guard-1.7bModel Category=Guard model (Prompt Injection), Number of Parameters=1.7b2026.05 | 56.57 | 17.2 | 80.88 | 28.37 | |
| DataSentinelModel Category=Guard model (Prompt Injection)2026.05 | 56.5 | 42.6 | 59 | 49.48 | |
| Prompt-Guard-2-86MModel Category=Guard model (Jail Break), Number of Parameters=86M2026.05 | 54.7 | 9.4 | 100 | 17.18 | |
| PromptArmorModel Category=Guard model (Prompt Injection)2026.05 | 54.33 | 53.33 | 54.42 | 53.87 | |
| Llama-Guard-3-Vision-11BModel Category=Guard model (Jail Break), Number of Parameters=11B2026.05 | 50 | 0 | 0 | 0 | |
| GuardReasoner-VL-7BModel Category=Guard model (Jail Break), Number of Parameters=7B2026.05 | 50 | 0 | 0 | 0 |