Jailbreak Attack on JBB (ASRpaper/ASRgen Metrics)
97ASR (Paper)TrojFill
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TrojFillVictim=GPT-4o, Judge=GPT-4o, Repo=Available, Gen. LOC=711, rho=1.84, Iter=2†, Search=Stateful, Carrier=Obfuscate2026.02 | 97 | 87.5 | -6.3 | |
| AIR*Victim=GPT-4o, Judge=LLaMA-3-70B, Repo=Available, Gen. LOC=487, rho=0.57, Iter=2†, Search=Single-pass, Carrier=Formal2026.02 | 95 | 100 | 5 | |
| JAIL-CON-CITVictim=GPT-4o, Judge=GPT-4o-mini, Repo=Available, Gen. LOC=211, rho=0.33, Iter=5, Search=Sampling, Carrier=Obfuscate2026.02 | 92 | 90.6 | -1.4 | |
| Past-TenseVictim=GPT-4o, Judge=GPT-4, Repo=Available, Gen. LOC=280, rho=0.88, Iter=2†, Search=Sampling, Carrier=Reframe2026.02 | 88 | 83.9 | -4.1 | |
| MousetrapVictim=Claude-3.5-Sonnet, Judge=GPT-4o, Repo=Available, Gen. LOC=462, rho=1.14, Iter=2, Search=Sampling, Carrier=Obfuscate2026.02 | 87 | 87.5 | 0.5 | |
| JAIL-CON-CVTVictim=GPT-4o, Judge=GPT-4o-mini, Repo=Available, Gen. LOC=211, rho=0.33, Iter=5, Search=Sampling, Carrier=Obfuscate2026.02 | 79 | 78.1 | -0.9 | |
| TRIAL*Victim=GPT-4o, Judge=Human(GPT-4o), Repo=None, Gen. LOC=417, Iter=5†, Search=Victim-loop, Carrier=Context2026.02 | 73 | 75 | 2 |