List your company
ResearchFormerly WizWand

Jailbreak Attack Success Rate on Harmful prompts dataset

97Attack Success RateDaredevil-8B
-3.8822.3148.574.69Feb 24, 2024Jun 21, 2024Oct 17, 2024Feb 12, 2025Jun 10, 2025Oct 6, 2025Feb 2, 2026
Updated 8mo ago

Evaluation results

MethodLinks
Daredevil-8BDecomposition Strategy=Monitor2026.0297
gpt-4o-miniDecomposition Strategy=Monitor2026.0296
Daredevil-8BDecomposition Strategy=Cost+Monitor2026.0292.5
Deepseek-v3.2Decomposition Strategy=Monitor2026.0292.3
gpt-4o-miniDecomposition Strategy=Cost+Monitor2026.0291
Daredevil-8BDecomposition Strategy=Cost2026.0289.5
PAPBase LLM=Vicuna-33B, Guard Model=None, Threat Model=No Access2024.0288
PRPBase LLM=Vicuna-33B, Guard Model=GPT3.5, Threat Model=No Access2024.0288
gemini-3-flashDecomposition Strategy=Cost+Monitor2026.0287.5
PAPBase LLM=Guanaco-13B, Guard Model=None, Threat Model=No Access2024.0284
PRPBase LLM=Guanaco-13B, Guard Model=GPT3.5, Threat Model=No Access2024.0284
Deepseek-v3.2Decomposition Strategy=Cost2026.0284
Daredevil-8BDecomposition Strategy=Standard2026.0283.5
Hermes3-70bDecomposition Strategy=Cost+Monitor2026.0282.5
Deepseek-v3.2Decomposition Strategy=Cost+Monitor2026.0281.2
PRPBase LLM=Llama2-70B, Guard Model=GPT3.5, Threat Model=No Access2024.0280
PRPBase LLM=Vicuna-33B, Guard Model=Llama2-70B, Threat Model=No Access2024.0280
PRPBase LLM=Vicuna-33B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0280
PRPBase LLM=Llama2-70B, Guard Model=Llama2-70B, Threat Model=No Access2024.0278
PRPBase LLM=Guanaco-13B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0278
Hermes3-70bDecomposition Strategy=Monitor2026.0277.8
PRPBase LLM=Guanaco-13B, Guard Model=Llama2-70B, Threat Model=No Access2024.0276
PRPBase LLM=Llama2-70B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0274
PAPBase LLM=Llama2-70B, Guard Model=None, Threat Model=No Access2024.0266
PAPBase LLM=Vicuna-33B, Guard Model=Llama2-70B, Threat Model=No Access2024.0264
PAPBase LLM=Llama2-70B, Guard Model=Llama2-70B, Threat Model=No Access2024.0256
PAPBase LLM=Vicuna-33B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0256
PAPBase LLM=Llama2-70B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0250
PAPBase LLM=Guanaco-13B, Guard Model=Llama2-70B, Threat Model=No Access2024.0250
gpt-4o-miniDecomposition Strategy=Cost2026.0249
PAPBase LLM=Guanaco-13B, Guard Model=Gemini-Pro, Threat Model=No Access2024.0246
Deepseek-v3.2Decomposition Strategy=Standard2026.0244.4
gemini-3-flashDecomposition Strategy=Cost2026.0237.5
Hermes3-70bDecomposition Strategy=Cost2026.0228
Hermes3-70bDecomposition Strategy=Standard2026.0226.5
gpt-4o-miniDecomposition Strategy=Standard2026.0219
PAPBase LLM=Vicuna-33B, Guard Model=GPT3.5, Threat Model=No Access2024.0212
claude-haiku-4.5Decomposition Strategy=Cost2026.027.1
PAPBase LLM=Guanaco-13B, Guard Model=GPT3.5, Threat Model=No Access2024.024
claude-haiku-4.5Decomposition Strategy=Cost+Monitor2026.023.7
gpt-oss-20bDecomposition Strategy=Cost+Monitor2026.023.5
claude-haiku-4.5Decomposition Strategy=Standard2026.021.1
gpt-oss-20bDecomposition Strategy=Standard2026.021
gpt-oss-20bDecomposition Strategy=Monitor2026.021
PAPBase LLM=Llama2-70B, Guard Model=GPT3.5, Threat Model=No Access2024.020
gpt-oss-20bDecomposition Strategy=Cost2026.020
gemini-3-flashDecomposition Strategy=Standard2026.020
gemini-3-flashDecomposition Strategy=Monitor2026.020
claude-haiku-4.5Decomposition Strategy=Monitor2026.020