ModelRadar
/
← All benchmarks

HarmBench

safety · jailbreak-robustness↗ Quelle
55Evidence
not yet measured

What it measures

Wie oft standardisierte Angriffe die Schutzmechanismen umgehen.

What it does not measure

Ueberblockierung harmloser Anfragen.

How the score is composed

Currency50
Discrimination50
Headroom50
Contamination safety33
Gaming resistance100
Methodological rigor80
Reproducibility60

Details

Benchmaxxing risk
low
Difficulty
medium
Public test set
yes
Human-verified
yes
LLM-judged
no
Skala
0–100 % attack success (niedriger besser)

Hand-maintained — no automatic adapter yet.

Best measured models

No results ingested yet.