This research reveals a significant vulnerability in text-to-image models that threatens the commercial prompt marketplace ecosystem.

• Introduces EvoStealer, a differential evolution approach that can extract valuable prompt templates using just a few image samples
• Creates Prism, a benchmark of 50 templates and 450 images to test prompt-stealing attacks
• Demonstrates that even complex, commercially valuable prompts can be reverse-engineered
• Highlights urgent security implications for AI artists, prompt engineers, and marketplace platforms

This work matters for security professionals as it exposes a novel IP theft vector in generative AI that could undermine emerging creative economies and requires technical countermeasures.

Vulnerability of Text-to-Image Models to Prompt Template Stealing: A Differential Evolution Approach