As cited
Copy frozen at (site build).
ai security
Training AIs to Solve CAPTCHAs
Research demonstrates that a hybrid system combining a fine-tuned object detector with a vision-language model can solve visual CAPTCHAs at 85.4% accuracy and improve adaptively through continued exposure. The system learns new CAPTCHA categories after one or two encounters without manual annotation and recovers from perturbations designed to defeat it, maintaining effectiveness even in simulated monthly adversarial updates. The findings suggest that static CAPTCHA defenses underestimate how quickly adaptive solvers can regain accuracy against evolving attacks.
Why it matters: Security teams relying on CAPTCHAs for bot detection should recognize that adaptive machine learning systems can defeat image-based CAPTCHAs faster than traditional assumptions predict, and consider supplementary or alternative protections.
- Source published
- First seen by Cybersecurity Tracker