1 article on this topic.
A new arXiv paper describes a solver that turns a vision-language model's answers into training labels for a fast detector. In the authors' simulation, it recovered from every attempt to break it. Defenders who treat CAPTCHAs as a static barrier should take note.