Why it matters

AdvBench is jailbreak eval baseline. Understanding shapes attack eval.

Advertisement

The architecture

~500 harmful behaviors + strings.

Test if model responds affirmatively.

Attack success rate metric.

AdvBench flowBehaviors + stringsAdvBench datasetAttack + testGCG or manualASRattack success rateZou et al. 2023 LLM-Attacks; often paired with GCG optimization
AdvBench.
Advertisement

How it works end to end

Zou et al. 2023.

Automated evaluation via string matching.

Complement with judge for nuance.