Computer-Using Agent Task on OS-Harm 1.0 (test)
0.99PCRSAFEPRED
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SAFEPREDFoundation Model=GPT-4o, Guardrail Model=Qwen3-Max2026.02 | 0.99 | 0.35 | 0.34 | — | — | — | — | — | — | — | — | — | |
| Generic-defenseFoundation Model=GPT-4o2026.02 | 0.95 | 0.16 | 0.15 | — | — | — | — | — | — | — | — | — | |
| HarmonyGuardFoundation Model=GPT-4o, Guardrail Model=Qwen3-Max2026.02 | 0.93 | 0.24 | 0.23 | — | — | — | — | — | — | — | — | — | |
| Rule-traversedFoundation Model=GPT-4o2026.02 | 0.92 | 0.17 | 0.14 | — | — | — | — | — | — | — | — | — | |
| NoneFoundation Model=GPT-4o2026.02 | 0.87 | 0.22 | 0.22 | — | — | — | — | — | — | — | — | — |