LLM Agent Security and Utility Evaluation on DTAP-150
1Attack Success Rate (ASR)ARGUS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ARGUSBackbone=Sonnet 4.62026.05 | 1 | 56 | 6 | |
| AIRGuardBackbone=Sonnet 4.62026.05 | 1 | 72 | 18 | |
| AIRGuardBackbone=GPT-5.3-codex2026.05 | 3 | 78 | 8 | |
| ARGUSBackbone=Haiku 4.52026.05 | 4 | 52 | 18 | |
| AIRGuardBackbone=GPT-5.4-mini2026.05 | 4 | 75 | 6 | |
| ARGUSBackbone=GPT-5.3-codex2026.05 | 4 | 60 | 0 | |
| MELONBackbone=Haiku 4.52026.05 | 5 | 42 | 32 | |
| MELONBackbone=GPT-5.4-mini2026.05 | 5 | 52 | 14 | |
| AIRGuardBackbone=Haiku 4.52026.05 | 6 | 76 | 18 | |
| MELONBackbone=GPT-5.3-codex2026.05 | 6 | 50 | 20 | |
| No defenseBackbone=Sonnet 4.62026.05 | 7 | 90 | 0 | |
| MELONBackbone=Sonnet 4.62026.05 | 7 | 42 | 28 | |
| ARGUSBackbone=GPT-5.4-mini2026.05 | 7 | 60 | 0 | |
| No defenseBackbone=Haiku 4.52026.05 | 18 | 94 | 0 | |
| No defenseBackbone=GPT-5.4-mini2026.05 | 22 | 86 | 0 | |
| No defenseBackbone=GPT-5.3-codex2026.05 | 27 | 90 | 0 |