Agent Safety Auditing on ASSE-Safety
85.4AccuracyTRACES
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TRACESBackbone=Llama3.1-8B2026.05 | 85.4 | 90.2 | 95.9 | 39.2 | 80.3 | |
| AgentDoGBackbone=Llama3.1-8B2026.05 | 83.4 | 82.1 | 84.3 | 26.8 | 64.9 | |
| AgentDoGBackbone=Qwen3-4B2026.05 | 82.8 | 83.8 | 81.5 | 18.5 | 59.7 | |
| TRACESBackbone=Qwen3-4B2026.05 | 81.6 | 87.5 | 92.6 | 34.5 | 76.9 | |
| ShieldAgent2026.05 | 77.8 | 80.2 | 91.5 | 15.5 | 59.8 | |
| AgentAuditor2026.05 | 74.4 | 77.2 | 74.6 | 22.1 | 67.3 | |
| GPT-5.22026.05 | 70.3 | 69.9 | 63 | 26.7 | 62.9 | |
| Gemini-3-Flash2026.05 | 61.1 | 46 | 30.2 | 2.7 | 62 | |
| PolyGuard2026.05 | 56.7 | 55.4 | 50.6 | 12.7 | 59.4 | |
| Qwen3-4B2026.05 | 54.9 | 70.9 | 100 | 61.6 | 50 | |
| Llama3.1-8B-Instruct2026.05 | 53.4 | 30.3 | 18.5 | 2.1 | 54.7 | |
| LlamaGuard3-8B2026.05 | 50.3 | 17.9 | 9.9 | 2.1 | 53.9 | |
| JoySafety2026.05 | 36.9 | 28.7 | 23.6 | 16 | 55.1 |