Attack Prevention on AGENTREDBENCH cross-integration held-out (test)
32.1Baseline ASRAGENTREDGUARD-MiniLM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AGENTREDGUARD-MiniLMTarget Model=Claude Sonnet 4.62026.06 | 32.1 | 0 | 32.1 | |
| AGENTREDGUARD-MiniLMTarget Model=GPT-5.4-nano2026.06 | 63.7 | 3.3 | 60.5 | |
| AGENTREDGUARD-MiniLMTarget Model=Overall2026.06 | 69.9 | 2.4 | 67.5 | |
| AGENTREDGUARD-MiniLMTarget Model=GPT-5.4-mini2026.06 | 71.6 | 1.9 | 69.8 | |
| AGENTREDGUARD-MiniLMTarget Model=GPT-5.42026.06 | 72.6 | 1.9 | 70.7 | |
| AGENTREDGUARD-MiniLMTarget Model=Gemini 3.1 Pro2026.06 | 78.6 | 3.3 | 75.3 | |
| AGENTREDGUARD-MiniLMTarget Model=Gemini 3.1 Flash Lite2026.06 | 79.5 | 3.3 | 76.3 | |
| AGENTREDGUARD-MiniLMTarget Model=Claude Haiku 4.52026.06 | 79.5 | 2.3 | 77.2 | |
| AGENTREDGUARD-MiniLMTarget Model=Gemini 3 Flash2026.06 | 81.4 | 3.3 | 78.1 |