Indirect Prompt Injection on InjecAgent
0.3Base ASRSanitizer
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SanitizerBackbone=GPT-4o2025.10 | 0.3 | 0 | |
| Prompt sandwichBackbone=GPT-4o2025.10 | 1 | 2 | |
| PI DetectorBackbone=GPT-4o2025.10 | 3.1 | 0 | |
| AgentWatcherAttribution LLM=Llama-3.1-8B-Instruct, Monitor LLM=Qwen-3-4B-Instruct-25072026.04 | 4 | 1 | |
| SpotlightingBackbone=GPT-4o2025.10 | 5.1 | 1.5 | |
| NoneBackbone=GPT-4o2025.10 | 8.3 | 3.8 | |
| GPT-OSS-SafeguardModel Version=20B2026.04 | 12 | 5 | |
| PromptGuard2026.04 | 13 | 2 | |
| PIGuard2026.04 | 15 | 3 | |
| PromptArmor2026.04 | 19 | 7 | |
| DataSentinel2026.04 | 21 | 25 | |
| No Defense2026.04 | 24 | 28 |