Adversarial Nibbler: Red-Teaming for Text-to-Image Safety Evaluation

Red-teaming methodology uncovers safety vulnerabilities in text-to-image models through implicitly adversarial prompts.