I work on bot detection at hCaptcha.com. (Hiring: reach out if you want to apply machine learning to stop bots and help websites monetize their traffic without

I work on bot detection at hCaptcha.com. (Hiring: reach out if you want to apply machine learning to stop bots and help websites monetize their traffic without ads) In order to successfully execute a replay attack you would also need to pass the Turing Test, i.e. click the correct images. If you design a bot that starts a combinatoric attacking by trying random guesses we can easily confuse it, so most attackers try to use a solver service. We can also identify how you interact with semantic content in the images when you click on the image and characterize your mouse interaction as human or non-human. Since confidence increases as more results come in we can also run them after the initial pass and then shadow-ban bots. (And notify the targeted website that we have determined e.g. a particular signup is a bot.) Ultimately, many techniques beyond simple correct/not-correct are required to defend against the main attack vector: humans hired to solve captchas en-masse and make thousands of fake accounts. Modern ML is pretty effective for these kinds of problems. Browser obfuscation does not add real security, and today's reCAPTCHA (all versions) is easily defeated in practice.