Guardrail Performance Evaluation on Scenario A Multi-Step Financial Transfer
100AccuracyGPT-5.2
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GPT-5.2Guardrail Paradigm=ePCA2026.05 | 100 | 0 | 100 | |
| GPT-5.4Guardrail Paradigm=ePCA2026.05 | 100 | 0 | 100 | |
| Qwen3-maxGuardrail Paradigm=LLM-as-the-Judge2026.05 | 100 | 0 | 100 | |
| Qwen3-maxGuardrail Paradigm=ePCA2026.05 | 100 | 0 | 100 | |
| Gemini-3-flashGuardrail Paradigm=ePCA2026.05 | 100 | 0 | 100 | |
| Kimi-k2.5Guardrail Paradigm=ePCA2026.05 | 100 | 0 | 100 | |
| Gemini-3-flashGuardrail Paradigm=ABAC2026.05 | 98.3 | 0 | 99.2 | |
| GPT-5.2Guardrail Paradigm=LLM-as-the-Judge2026.05 | 83.3 | 0 | 90.9 | |
| Gemini-3-flashGuardrail Paradigm=LLM-as-the-Judge2026.05 | 75 | 0 | 85.7 | |
| GPT-5.4Guardrail Paradigm=LLM-as-the-Judge2026.05 | 53.3 | 6.7 | 68.1 | |
| GPT-5.2Guardrail Paradigm=ABAC2026.05 | 50 | 0 | 66.7 | |
| GPT-5.4Guardrail Paradigm=ABAC2026.05 | 50 | 0 | 66.7 | |
| Qwen3-maxGuardrail Paradigm=ABAC2026.05 | 50 | 0 | 66.7 | |
| Kimi-k2.5Guardrail Paradigm=ABAC2026.05 | 50 | 0 | 66.7 | |
| Kimi-k2.5Guardrail Paradigm=LLM-as-the-Judge2026.05 | 43.3 | 0 | 60.5 |