RC RANDOM CHAOS

Self-Improving AI Defeats CAPTCHA Defenses

· via Schneier on Security

Original source

Training AIs to Solve CAPTCHAs

Schneier on Security →

Vision-language models can solve visual CAPTCHAs but approach each puzzle independently, incurring identical computational cost regardless of repetition. Specialized detectors work in milliseconds but only for categories they were trained on, and neither approach improves from exposure.

Researchers combined a fine-tuned YOLOv8 detector with an open-weight VLM through a confidence-based router, running solely from screenshots and system input without DOM access. Every prediction from the VLM automatically labels training data, allowing the detector to absorb new CAPTCHA categories typically after one or two encounters. This feedback mechanism also enables recovery when CAPTCHA operators perturb images to disable the detector.

The system achieved 85.4% overall and 84.2% macro accuracy across 16 classes, surpassing either component alone. In a year-long simulated adversarial scenario with monthly perturbations, the solver recovered each round—and proved resilient even when trained by only a ~70%-accurate teacher model. The findings challenge conventional CAPTCHA defense assumptions about how quickly adaptive solvers can bounce back from sabotage.

Read the full article

Continue reading at Schneier on Security →

This is an AI-generated summary. Read the original for the full story.