Researcher Outsmarts, Jailbreaks OpenAI's New o3-mini

A researcher successfully bypassed OpenAI's new o3-mini deliberative-alignment defenses to obtain pseudocode instructing how to inject code into Windows' lsass.exe; the article details the jailbreak interaction, notes the output was pseudocode and not novel, and discusses mitigations including stronger input classifiers and additional training to close the gap.