Multi-turn Jailbreak Detection on HarmBench and DEFCON Multi-turn Jailbreak N=1,010 (test)
84F1 ScoreDeepContext
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DeepContext2026.02 | 84 | 83 | 86 | 4.24 | |
| Llama-Prompt-Guard-2Parameters=86M2026.02 | 67 | 60 | 76 | 5.83 | |
| Granite-Guardian-3.3Parameters=8B2026.02 | 67 | 57 | 83 | 5.03 | |
| Gpt5-Nano2026.02 | 65 | 55 | 81 | 5.73 | |
| Deberta-v3-Prompt-InjectionBackbone=Deberta-v32026.02 | 62 | 61 | 62 | 5.27 | |
| GCP Model Armor2026.02 | 58 | 54 | 63 | 4.56 | |
| Qwen3Guard-GenParameters=8B2026.02 | 51 | 36 | 84 | 3.12 | |
| Llama-Guard-4Parameters=12B2026.02 | 51 | 42 | 65 | 6.09 | |
| AWS Prompt Attack Guardrails2026.02 | 38 | 40 | 36 | 5.61 | |
| Azure Prompt Shield2026.02 | 19 | 11 | 62 | 8 |