AI Solver Beats CAPTCHA Defenses

Researchers demonstrate a vision-language model system that solves visual CAPTCHAs with 85.4% accuracy, surpassing specialized detectors. The system uses a fine-tuned YOLOv8 detector paired with an open-weight VLM to learn from screenshots, recovering from adversarial perturbations within one or two encounters without human annotation during a simulated year-long arms race.