Best-of-N Jailbreaking (BoN Jailbreaking)

A simple black-box attack that tries many variants of the same prompt, applying augmentations such as random word shuffling or changes in capitalization, until it elicits a harmful response. Work by Hughes, Price et al. (December 2024) showed a success rate of around 89% on GPT-4o and 78% on Claude 3.5 Sonnet.