List your company
ResearchFormerly WizWand

Jailbreak Attack on HarmBench example-based Llama3 8B

5Attack Success RateCoA
1.226.8552.578.15Aug 1, 2025Aug 24, 2025Sep 16, 2025Oct 9, 2025Nov 1, 2025Nov 24, 2025Dec 18, 2025
Updated 6mo ago

Evaluation results

MethodLinks
CoAEvaluation Judge=GPT-Judge2025.0851.5
AmpleGCGEvaluation Judge=GPT-Judge2025.0861.39
I-GCGEvaluation Judge=GPT-Judge2025.0871.5
PiFEvaluation Judge=GPT-Judge2025.088.52.37
PAIREvaluation Judge=GPT-Judge2025.08152.93
ActorBreakerEvaluation Judge=GPT-Judge2025.08233.94
CrescendoEvaluation Judge=GPT-Judge2025.0825.53.3
GCGEvaluation Judge=GPT-Judge2025.08351.44
ReNeLLMEvaluation Judge=GPT-Judge2025.0843.54.12
GCG AttackHuman Readable=false, Zero-shot=true2025.1244—
AutoDAN TurboHuman Readable=true, Zero-shot=true2025.1262—
PAIR AttackHuman Readable=true, Zero-shot=true2025.1266—
AutoDANEvaluation Judge=GPT-Judge2025.0868.54.47
AGILEGenerator LLM=Llama-3-8B-Instruct, p=5, NCand=25, Evaluation Judge=GPT-Judge2025.08764.67
Adversarial ReasoningHuman Readable=true, Zero-shot=true2025.1288—
Adaptive AttackHuman Readable=false, Zero-shot=true2025.12100—
Jailbreak-ZeroHuman Readable=true, Zero-shot=true2025.12100—