Offensive-security instruction prompts for refusal-direction research and abliteration of code/security models.
This dataset contains infosec-domain harmful prompts intended to elicit refusal behavior from aligned instruction models. It is designed as the harmful side of a harmful/harmless contrast pair, analogous to mlabonne/harmful_behaviors but focused on offensive-security and malicious-coding requests.
Rows:
