CLIP and OWLv2 Running Locally Break Google reCAPTCHA at a 92.6% Per-Session Rate and Zero Cost

arXiv 2609.18518·medium signalXBlueskyLinkedInCopy link

arXiv 2609.18518 (16 Sep 2026) replaces the paid cloud VLM assumed by prior CAPTCHA-breaking work with free local open-source models. The authors taxonomize reCAPTCHA into Type A (independent image tiles, static and dynamic) and Type B (one image split into a 4x4 grid), then build zero-shot untrained solvers reaching 58% per-challenge accuracy on Type A with CLIP and 43.5% on Type B with OWLv2. End to end the automated solver succeeds in 92.6% of 500 real-world reCAPTCHA sessions with no API access and no per-attempt cost, which removes the economic friction that kept the defense viable.