Prompt Injection Detection on NeuralExec (test)
98.8Detection Accuracy (Top-1)CONTEXTCITE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CONTEXTCITEModel=Llama3-8B, Attack Type=Optimization-based2024.09 | 98.8 | 100 | |
| CONTEXTCITEModel=Phi-3-mini, Attack Type=Handcrafted2024.09 | 90 | 100 | |
| CONTEXTCITEModel=Phi-3-mini, Attack Type=Optimization-based2024.09 | 85 | 85 |