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

A new research paper introduces PersonaTeaming, a framework that integrates human-like personas into automated red-teaming for generative AI systems. Unlike traditional automated methods that generate adversarial prompts without considering the tester's background or perspective, PersonaTeaming creates personas — simulated user identities with specific traits, biases, and goals — to guide prompt generation. The core workflow systematically incorporates these personas into the adversarial prompt pipeline, exploring a broader range of attack strategies than prior methods like RainbowPlus. According to the paper, PersonaTeaming Workflow achieves higher attack success rates while preserving prompt diversity, a critical balance often hard to reach in automated safety testing.

To bridge the gap between automated persona simulation and real human judgment, the team also built PersonaTeaming Playground, an interactive interface that allows human red-teamers to author their own personas and co-create adversarial prompts with an AI assistant. In a user study with 11 industry practitioners, participants found the tool produced diverse, useful outputs and reported that AI-generated suggestions sparked creative thinking even when not strictly followed. The work advances both fully automated and human-in-the-loop red-teaming approaches, offering design insights for safer generative AI deployment.