Jailbreak Attack on JBDistill-Bench
1.2Attack Success Rate (ASR)GPT-5.2
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-5.2Model Category=Frontier, Evaluation Protocol=Three-judge majority vote, Development Status=Held-out2026.04 | 1.2 | |
| GPT-5.4*Model Category=Frontier, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Frontier dev target2026.04 | 1.4 | |
| Grok 4Model Category=Frontier, Evaluation Protocol=Three-judge majority vote, Development Status=Held-out2026.04 | 10.6 | |
| o4-mini*Model Category=Frontier, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Frontier dev target2026.04 | 10.6 | |
| Kimi-K2.5Model Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=Held-out2026.04 | 24.6 | |
| GPT-4o†Model Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Standard dev target2026.04 | 36.2 | |
| GPT-4.1-miniModel Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=Held-out2026.04 | 37 | |
| DeepSeek-V3.2†Model Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Standard dev target2026.04 | 39.2 | |
| GPT-4.1†Model Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Standard dev target2026.04 | 39.6 | |
| Phi-4-miniModel Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=Held-out2026.04 | 41.6 | |
| Grok 3†Model Category=Standard, Evaluation Protocol=Three-judge majority vote, Development Status=AR-Standard dev target2026.04 | 46.8 |