Adaptive AI solver learns new CAPTCHA categories from its own answers

Researchers report that an adaptive AI system solved visual CAPTCHAs with 85.4% overall accuracy across 16 classes, according to research highlighted by Schneier on Security. The system routes screenshots between a fine-tuned YOLOv8 detector and an open-weight vision-language model based on confidence, using operating-system input events without browser automation or access to page structure. Answers from the vision-language model become training labels, allowing the detector to learn new categories, typically after one or two encounters without human annotation. Image perturbations drove the detector's accuracy to 0%, but labels from the unaffected vision-language model enabled recovery. In a year-long simulated contest with defenses revised monthly, the solver recovered every round.