CAPTCHAs are widely deployed as human verification mechanisms and frequently block intelligent agents from completing end-to-end automation in real-world web environments. Existing CAPTCHA-solving methods rely on prompting closed-source models, incurring high inference costs and lacking controllability, yet training-based approaches have remained absent due to the lack of large-scale training data with process-level annotations. To enable training-based CAPTCHA solving, we introduce CaptchaBench, a large-scale benchmark of 16,000 programmatically generated samples with detailed region and process-level annotations. Systematic evaluation on CaptchaBench reveals that existing methods fail on tasks requiring fine-grained visual detail capture, with a key challenge being the inability to supervise the model’s reasoning process over task-relevant visual regions. Therefore, we present CaptchaMind, an RL-based solver that explicitly supervises the model’s reasoning process by directly rewarding correct identification of task-relevant visual regions at intermediate steps. CaptchaMind achieves state-of-the-art performance across eight tasks and generalizes effectively to real-world CAPTCHA instances, substantially outperforming all existing methods without closed-source APIs.