Guard Detection on AGENTREDBENCH (cross-integration held-out full benign corpus)
99.9TPRAGENTREDGUARD-DeBERTa
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AGENTREDGUARD-DeBERTaParams=142M, Threshold=0.52026.06 | 99.9 | 0.29 | |
| AGENTREDGUARD-MiniLMParams=23M, Threshold=0.52026.06 | 99.75 | 0.37 | |
| ProtectAI PIParams=184M, Threshold=0.5, Backbone=deberta-v3-base-prompt-injection-v22026.06 | 15.03 | 25.05 | |
| WildGuardParams=7B, Threshold=0.52026.06 | 0.25 | 0.12 | |
| Llama Guard 3Params=8B, Threshold=0.52026.06 | 0.15 | 2.39 | |
| PromptGuard 2Params=22M, Threshold=0.52026.06 | 0 | 1.93 |