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

PersonaTeaming introduces a persona-driven approach to red-teaming generative AI systems, combining automated adversarial prompt generation with human-in-the-loop collaboration. The method outperforms existing automated approaches while enabling security researchers to leverage diverse perspectives and backgrounds to uncover AI model vulnerabilities more effectively.

PersonaTeaming addresses a critical gap in AI safety research by recognizing that red-teaming effectiveness depends heavily on the testers' backgrounds, identities, and perspectives. Traditional automated red-teaming methods treat adversarial prompt generation as a uniform process, missing the insight that different human experiences surface different risks. This research validates what security practitioners intuitively understand: threat discovery requires diverse thinking patterns.

The dual-approach framework—combining PersonaTeaming Workflow for scalable automated testing with PersonaTeaming Playground for collaborative human-AI interaction—represents a pragmatic advancement in AI governance. The workflow achieves higher attack success rates than RainbowPlus, a leading baseline, while the playground interface demonstrates measurable value in user studies with industry practitioners. Importantly, the research reveals that AI-generated suggestions enhance human creativity even when not directly adopted, suggesting productive complementarity between human judgment and machine capabilities.

For the AI safety ecosystem, this work has immediate practical implications. Organizations building or deploying generative AI systems need more sophisticated red-teaming capabilities to identify edge cases and harmful behaviors before release. PersonaTeaming's emphasis on persona-driven exploration could become standard practice in responsible AI development pipelines. The research also informs the broader debate about human-AI collaboration, demonstrating that effective partnerships require systems designed around human agency rather than full automation.

The trajectory suggests future red-teaming tools will increasingly incorporate demographic, cultural, and experiential diversity as explicit design features. As regulatory frameworks for AI safety tighten globally, methodologies like PersonaTeaming that systematically explore vulnerability landscapes become competitive advantages for responsible developers.

• →PersonaTeaming achieves higher attack success rates than state-of-the-art automated methods while maintaining prompt diversity through persona-driven prompt generation.
• →Human-in-the-loop design enables red-teamers to author custom personas and collaborate with AI to refine adversarial prompts, producing outputs practitioners find more useful.
• →Diverse perspectives and backgrounds fundamentally shape which AI vulnerabilities get discovered, supporting the need for persona-inclusive red-teaming approaches.
• →AI-generated suggestions enhance creative thinking patterns even when practitioners do not directly follow them, demonstrating productive human-AI collaboration dynamics.
• →Persona-driven red-teaming represents an emerging best practice for organizations developing generative AI systems to identify and mitigate safety risks before deployment.