Attack Detection on Agent Traces (Held-out)
0.89ROC-AUC (Social engineering)gated multi-view fusion
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| gated multi-view fusionRepresentation=Gated2026.01 | 0.89 | 0.83 | 0.62 | 0.6 | 0.92 | |
| Conversational tokenizationRepresentation=Conversational2026.01 | 0.78 | 0.69 | 0.46 | 0.39 | 0.26 | |
| structural tokenizationRepresentation=Structural2026.01 | 0.67 | 0.81 | 0.85 | 0.85 | 0.97 |