Prompt Injection Defense on LLMail full 22,899-attack pool
0ASRPipeline
Evaluation results
| Method | Links | |
|---|---|---|
| PipelineModel=gpt-5-mini2026.03 | 0 | |
| Two-AgentModel=gpt-5-mini2026.03 | 0.009 | |
| JSON ValidatorModel=gpt-5-mini2026.03 | 0.4 | |
| All combinedModel=GPT-4o mini2026.03 | 2 | |
| BaselineModel=gpt-5-mini2026.03 | 2.83 | |
| TaskTrackerModel=GPT-4o mini2026.03 | 5 | |
| LLM-as-judgeModel=GPT-4o mini2026.03 | 8 | |
| PromptShieldModel=GPT-4o mini2026.03 | 10 | |
| SpotlightingModel=GPT-4o mini2026.03 | 15 | |
| None (original)Model=Phi-3 / GPT-4o mini2026.03 | 30 |