List your company

Malicious Prompt Refusal on HarmBench

96Refusal Ratemicrosoft/Phi-3-mini-128k-instruct
16.9637.485878.52Dec 12, 2025Jan 12, 2026Feb 13, 2026Mar 17, 2026Apr 17, 2026May 19, 2026Jun 20, 2026

Evaluation results

MethodLinks
microsoft/Phi-3-mini-128k-instructSuffix Type=No Suffix2025.1296—
Phi-3-mini-128k-InstructSuffix Type=No Suffix2026.0696—
google/gemma-2b-itSuffix Type=No Suffix2025.1294—
Gemma-2B-ITSuffix Type=No Suffix2026.0694—
meta-llama/Llama-3.1-8B-instructSuffix Type=No Suffix2025.1289—
Llama-3.1-8B-InstructSuffix Type=No Suffix2026.0689—
meta-llama/Llama-3.2-3B-instructSuffix Type=No Suffix2025.1284—
Llama-3.2-3B-InstructSuffix Type=No Suffix2026.0684—
SafeAnchorBase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0478.4—
lmsys/vicuna-7b-v1.5Suffix Type=Primary Suffix2025.12760.32
Vicuna-7B-v1.5Suffix Type=Primary Suffix2026.067632
google/gemma-2b-itSuffix Type=Super Suffix2025.12620.91
Gemma-2B-ITSuffix Type=Super Suffix2026.066291
microsoft/Phi-3-mini-128k-instructSuffix Type=Super Suffix2025.12590.86
Phi-3-mini-128k-InstructSuffix Type=Super Suffix2026.065986
lmsys/vicuna-7b-v1.5Suffix Type=Super Suffix2025.12580.86
Vicuna-7B-v1.5Suffix Type=Super Suffix2026.065886
meta-llama/Llama-3.1-8B-instructSuffix Type=Super Suffix2025.12550.88
meta-llama/Llama-3.2-3B-instructSuffix Type=Super Suffix2025.12550.9
Llama-3.1-8B-InstructSuffix Type=Super Suffix2026.065588
Llama-3.2-3B-InstructSuffix Type=Super Suffix2026.065590
SafeGrad + LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0454.6—
meta-llama/Llama-3.2-3B-instructSuffix Type=Primary Suffix2025.12520.2
Llama-3.2-3B-InstructSuffix Type=Primary Suffix2026.065220
google/gemma-2b-itSuffix Type=Primary Suffix2025.12510.35
lmsys/vicuna-7b-v1.5Suffix Type=No Suffix2025.1251—
microsoft/Phi-3-mini-128k-instructSuffix Type=Primary Suffix2025.12510.17
Gemma-2B-ITSuffix Type=Primary Suffix2026.065135
Vicuna-7B-v1.5Suffix Type=No Suffix2026.0651—
Phi-3-mini-128k-InstructSuffix Type=Primary Suffix2026.065117
Safety InterleavingBase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0449.3—
Vaccine + LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0446.5—
Safe LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0444.7—
EWC + LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0436.4—
O-LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0434.8—
Standard LoRABase Model=Llama-2-7B-Chat, Attack Method=GCG, Adaptation Pipeline=default-ordering T=32026.0431.2—
meta-llama/Llama-3.1-8B-instructSuffix Type=Primary Suffix2025.12200.23
Llama-3.1-8B-InstructSuffix Type=Primary Suffix2026.062023