CAPTCHA Hardening Lab
Security ResearchHow easily a vision-language model reads a CAPTCHA, and what hardening actually defeats it
Upload a CAPTCHA-style image (a distorted-text challenge you already have — this only reads an image you provide, it never contacts or solves a live CAPTCHA on a real website). A vision-language model attempts to read it, then a hardening slider stacks three classic, model-agnostic distortions — pixel noise, an occlusion wave, and reduced contrast/color — and the model tries again on the hardened version. The real question is how much hardening it takes before the model's answer breaks, not whether this one attempt succeeds.