Direct Prompt Injection robustness on Agent Security Bench DPI
19ASRPhi-4
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Phi-4Safety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 19 | 81 | |
| GPT-4oSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 21 | 79 | |
| GPT-5Safety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 26 | 67 | |
| Phi-4Safety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 28 | 72 | |
| Qwen3-4B-ThinkSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 29 | 71 | |
| GPT-5Safety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 42 | 0 | |
| Qwen2.5-7BSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 42 | 58 | |
| Qwen3-4B-ThinkSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 46 | 46 | |
| Qwen2.5-7BSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 55 | 42 | |
| GPT-4oSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 76 | 0 |