Jailbreak Attack on JBB-Behaviors (Specific Robustness Metrics)
100S2C (Ours)Structured Semantic Cloaking
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Structured Semantic CloakingFamily=DeepSeek, Target Model=DeepSeek-r1, Type=Reasoning2026.03 | 100 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=DeepSeek, Target Model=DeepSeek-v3, Type=Instruction2026.03 | 100 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=GPT, Target Model=GPT-4.1-mini, Type=Instruction2026.03 | 100 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Gemini, Target Model=Gemma3-27b-it, Type=Instruction2026.03 | 100 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=DeepSeek, Target Model=DeepSeek-v3.2, Type=Reasoning2026.03 | 99 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Gemini, Target Model=Gemini-2.5, Type=Reasoning2026.03 | 99 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Qwen, Target Model=Qwen2.5-72B-Instruct, Type=Instruction2026.03 | 96 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Qwen, Target Model=Qwen3-32B, Type=Reasoning2026.03 | 95 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=GPT, Target Model=GPT-4o, Type=Instruction2026.03 | 95 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Llama, Target Model=Llama 3.3-70B-Instruct, Type=Instruction2026.03 | 94 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=GPT, Target Model=GPT-4o-mini, Type=Instruction2026.03 | 90 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Claude, Target Model=Claude-3-7-Sonnet, Type=Instruction2026.03 | 86 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=GPT, Target Model=GPT-OSS-20B, Type=Reasoning2026.03 | 82 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Phi, Target Model=Phi-4, Type=Instruction2026.03 | 82 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=Claude, Target Model=Claude-4.5-Haiku, Type=Instruction2026.03 | 63 | — | — | — | — | — | — | — | |
| Structured Semantic CloakingFamily=GPT, Target Model=GPT-5-mini, Type=Reasoning2026.03 | 50 | — | — | — | — | — | — | — | |
| CipherFamily=Claude, Target Model=Claude-4.5-Haiku, Type=Instruction2026.03 | — | — | — | — | — | 0 | — | — | |
| CodeChameleonFamily=Llama, Target Model=Llama 3.3-70B-Instruct, Type=Instruction2026.03 | — | 91 | — | — | — | — | — | — | |
| HaPLaFamily=Gemini, Target Model=Gemini-2.5, Type=Reasoning2026.03 | — | — | — | — | — | — | 98 | — | |
| PAIRFamily=GPT, Target Model=GPT-4o, Type=Instruction2026.03 | — | — | — | — | 54 | — | — | — | |
| PAPFamily=Phi, Target Model=Phi-4, Type=Instruction2026.03 | — | — | — | — | — | — | — | 36 | |
| ReNeLLMFamily=DeepSeek, Target Model=DeepSeek-r1, Type=Reasoning2026.03 | — | — | — | 44 | — | — | — | — | |
| Word-based PuzzleFamily=Llama, Target Model=Llama 3.3-70B-Instruct, Type=Instruction2026.03 | — | — | 76 | — | — | — | — | — |