Skip to content
arXiv· arXiv:2605.05682· Wesley Hanwen Deng, Mingxi Yan, Sunnie S. Y. Kim, Akshita Jha, Lauren Wilcox, Kenneth Holstein, Motahhare Eslami, Leon A. Gatys·Indexed by this site · Originally published

PersonaTeaming: Supporting Persona-Driven Red Teaming for Generative AI

PersonaTeaming: Supporting Persona-Driven Red-Teaming for Generative AI

AI summary

The researchers propose the PersonaTeaming method, which introduces persona perspectives into the adversarial prompt generation process of generative AI red teaming to explore a broader range of adversarial strategies. Its automated workflow achieves a higher attack success rate than the state-of-the-art automated red teaming method RainbowPlus while maintaining prompt diversity. Because automatically generated personas can only approximate real human perspectives, the team further implements it as a user-facing interface, PersonaTeaming Playground, allowing red teamers to build their own personas and collaborate with AI to mutate and improve prompts. In a user study with 11 industry practitioners, the tool supported diverse red teaming strategies and outputs that practitioners found useful; AI-generated suggestions, even when not strictly followed, promoted out-of-the-box thinking.

Read originalarxiv.org