Safety-Utility Trade-off Evaluation on AgentHarm
82.79HS ScoreTRIAD + Tri-Guard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TRIAD + Tri-GuardBackbone=Kimi-2.5, Agent Category=Open-Source Agents2026.06 | 82.79 | 13.26 | |
| TRIAD + TS-GuardBackbone=Kimi-2.5, Agent Category=Open-Source Agents2026.06 | 82.25 | 15.32 | |
| TRIAD + Tri-GuardBackbone=Gemini-2.5-Pro, Agent Category=Proprietary Agents2026.06 | 80.85 | 15.39 | |
| TRIAD + Tri-GuardBackbone=Qwen3-32B, Agent Category=Open-Source Agents2026.06 | 80.32 | 9.87 | |
| ReActBackbone=GPT-5.1, Agent Category=Proprietary Agents2026.06 | 80.31 | 21.71 | |
| TRIAD + Tri-GuardBackbone=GPT-5.1, Agent Category=Proprietary Agents2026.06 | 79.73 | 13.66 | |
| TRIAD + TS-GuardBackbone=GPT-5.1, Agent Category=Proprietary Agents2026.06 | 77.58 | 15.08 | |
| TRIAD + TS-GuardBackbone=Gemini-2.5-Pro, Agent Category=Proprietary Agents2026.06 | 76.72 | 14.81 | |
| TRIAD + TS-GuardBackbone=Qwen3-32B, Agent Category=Open-Source Agents2026.06 | 76.15 | 8.89 | |
| ReActBackbone=Kimi-2.5, Agent Category=Open-Source Agents2026.06 | 74.57 | 36.9 | |
| ToolSafeBackbone=Kimi-2.5, Agent Category=Open-Source Agents2026.06 | 68.71 | 6.59 | |
| ToolSafeBackbone=Gemini-2.5-Pro, Agent Category=Proprietary Agents2026.06 | 68.49 | 10.38 | |
| ToolSafeBackbone=Qwen3-32B, Agent Category=Open-Source Agents2026.06 | 67.31 | 11.57 | |
| ReActBackbone=Gemini-2.5-Pro, Agent Category=Proprietary Agents2026.06 | 58.54 | 55.67 | |
| ReActBackbone=Qwen3-32B, Agent Category=Open-Source Agents2026.06 | 36.28 | 77.04 |