Benchmarking Leading AI Agents Against CAPTCHAs
research.roundtable.ai
We evaluate three leading AI models—Claude Sonnet 4.5, Gemini 2.5 Pro, and GPT-5—on their ability to solve Google reCAPTCHA v2 challenges and find significant performance differences, with success rates ranging from 28% to 60%.
  • finitebanjo@piefed.worldEnglish
    2·
    24 days ago

    Technically this very task is an act of data poisoning as those captchas are used to train detection algorithms.