Safety
h4rm3l
h4rm3l composable-jailbreak results: per-(target model, attack program, AdvBench prompt) binary success from a GPT-4 "BAD BOT" classifier. 6 target models x 113 synthesized composable attack programs x 50 AdvBench harmful prompts. response=1 when the attack jailbroke the target (BAD BOT), 0 when it refused (GOOD BOT); UNCLEAR/empty verdicts are dropped.
48items
6subjects
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
- 1gpt-3.5-turbo0.7551
- 2gpt-4o-2024-05-130.6388
- 3claude-3-haiku-202403070.2686
- 4meta-llama/Meta-Llama-3-70B-Instruct0.2673
- 5meta-llama/Meta-Llama-3-8B-Instruct0.1667
- 6claude-3-sonnet-202402290.106