Jailbreaking on AdvBench deepseek (520 malicious intent prompts)
0.4ASRprompt-injection
Evaluation Results
| Method | Links | |
|---|---|---|
| prompt-injectionAbbr.=pi, Pipeline=Single-turn baseline2026.05 | 0.4 | |
| obfuscationAbbr.=obs, Pipeline=Single-turn baseline2026.05 | 0.4 | |
| paraphrasingAbbr.=pp, Pipeline=Single-turn baseline2026.05 | 2.9 | |
| encryptionAbbr.=enc, Pipeline=Single-turn baseline2026.05 | 4 | |
| translationAbbr.=tr, Pipeline=Single-turn baseline2026.05 | 6.2 | |
| forced-completionAbbr.=fc, Pipeline=Single-turn baseline2026.05 | 6.7 | |
| nshot-hackingAbbr.=ns, Pipeline=Single-turn baseline2026.05 | 7.7 | |
| roleplayAbbr.=rp, Pipeline=Single-turn baseline2026.05 | 8.3 | |
| privilege-escalationAbbr.=pe, Pipeline=Single-turn baseline2026.05 | 12.9 | |
| cognitive-hackingAbbr.=ch, Pipeline=Single-turn baseline2026.05 | 14.4 | |
| gaslightingAbbr.=gas, Pipeline=Single-turn baseline2026.05 | 16.3 | |
| fictionalAbbr.=fic, Pipeline=Single-turn baseline2026.05 | 63.1 |