Indirect Prompt Injection Robustness on Agent Security Bench IPI
2Attack Success Rate (ASR)GPT-5
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GPT-5Safety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 2 | 65 | |
| GPT-5Safety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 3 | 0 | |
| GPT-4oSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 27 | 63 | |
| Phi-4Safety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 28 | 68 | |
| Qwen2.5-7BSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 33 | 61 | |
| Phi-4Safety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 39 | 54 | |
| Qwen2.5-7BSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 40 | 44 | |
| Qwen3-4B-ThinkSafety Scaffolding=MOSAIC, Evaluation Mode=Zero-shot2026.03 | 43 | 42 | |
| Qwen3-4B-ThinkSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 46 | 31 | |
| GPT-4oSafety Scaffolding=Base, Evaluation Mode=Zero-shot2026.03 | 67 | 0 |