RESEARCHER OUTSMARTS, JAILBREAKS OPENAI'S NEW O3-MINI

Summary from the TLDR InfoSec newsletter

A security researcher, Eran Shimony has successfully bypassed OpenAI's new "deliberative alignment" safety feature in o3-mini, getting it to provide Windows exploit instructions by posing as a historian. According to Shimony, OpenAI's models are particularly susceptible to natural language social engineering attacks, while Meta's Llama models are instead vulnerable to ASCII art-encoded malicious prompts, and Claude is most vulnerable to generating harmful code because it doesn't effectively classify code's potential malicious uses.