Human-crafted jailbreak defense on mhj
14Attack Success Rate (ASR)Cognitive Firewall
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Cognitive FirewallDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 14 | 86 | |
| Granite Guardian 8BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 26 | 66 | |
| Llama Guard 3 8BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 38 | 54 | |
| THRD-styleDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 40 | 50 | |
| ShieldGemma 9BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 44 | 40 | |
| TCA-styleDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 50 | 40 | |
| none (undefended)Defense Type=undefended, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 74 | 0 |