Safety
JailbreakBench
JailbreakBench artifact verdicts: per-(model, behavior, attack-method) jailbroken booleans from the JailbreakBench judge. Six (attack_method, attack_type) conditions across 4 target models and 100 harmful behaviors. Sparse: white-box attacks (DSN, GCG/white_box) only cover open-weight models.
105items
4subjects
MITlicense
safetydomain
textmodality
item-level responses released
Saturation status: Unknown
Response matrix
Fit to width. Hover for subject & item; click a cell for details.
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect
Subjects
- 1vicuna-13b-v1.50.846
- 2gpt-3.5-turbo-11060.5275
- 3llama-2-7b-chat-hf0.374
- 4gpt-4-0125-preview0.29