Politely Worded Jailbreaks: Tell GPT-4o the Right Keywords and Its Manners Fall Apart
Ourtake
Turns out the bias filters were the easiest door in the building.
What happened
Researchers at Theori Inc published a study showing that jailbreak prompts containing demographic keywords for marginalized groups succeeded more often against models such as GPT-4o. They introduced the PCJailbreak method and a PCDefense method that reduces the gap, and released the code as open source.