LLM Agent Security and Utility Evaluation on AgentTrap
3.3Attack Success Rate (ASR)ARGUS
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ARGUSBackbone=Haiku 4.52026.05 | 3.3 | 81.5 | 2 | |
| AIRGuardBackbone=Haiku 4.52026.05 | 3.3 | 84 | 4 | |
| ARGUSBackbone=Sonnet 4.62026.05 | 5.5 | 88 | 2 | |
| AIRGuardBackbone=Sonnet 4.62026.05 | 5.5 | 84 | 4 | |
| AIRGuardBackbone=GPT-5.4-mini2026.05 | 8.8 | 90 | 2 | |
| AIRGuardBackbone=GPT-5.3-codex2026.05 | 9.9 | 94 | 4 | |
| MELONBackbone=Haiku 4.52026.05 | 11 | 86 | 4 | |
| No defenseBackbone=GPT-5.4-mini2026.05 | 13.2 | 94 | 0 | |
| MELONBackbone=GPT-5.4-mini2026.05 | 18.7 | 96 | 0 | |
| No defenseBackbone=Haiku 4.52026.05 | 20.9 | 88 | 0 | |
| ARGUSBackbone=GPT-5.4-mini2026.05 | 22 | 91.5 | 2 | |
| MELONBackbone=Sonnet 4.62026.05 | 25.3 | 92 | 0 | |
| ARGUSBackbone=GPT-5.3-codex2026.05 | 25.3 | 91.7 | 6 | |
| MELONBackbone=GPT-5.3-codex2026.05 | 26.4 | 94 | 0 | |
| No defenseBackbone=Sonnet 4.62026.05 | 36.3 | 88 | 4 | |
| No defenseBackbone=GPT-5.3-codex2026.05 | 63.7 | 94 | 0 |