AI cracks reCAPTCHA: New challenges for online security
The CAPTCHA security system has been cracked by artificial intelligence. As demonstrated by researchers from ETH Zurich, a well-prepared model is capable of solving security "puzzles" in such a way that the system cannot recognize that the solver is not a human. The issue concerns the reCAPTCHAv2 variant.
Tech Radar describes the discovery based on an analysis shared by the researchers. In the described cases, a popular AI model called YOLO was used, which "on behalf of a human" solved reCAPTCHAv2 tasks—a system designed to effectively distinguish machines from humans. Well-known to everyone, these puzzles involve indicating images with specific content, such as all those showing traffic lights or motorcycles.
Until now, such a test was generally considered an effective method for verifying whether a human is genuinely at the computer, or whether the task is being performed by some type of script. However, as the described studies show, a well-prepared YOLO model (trained here based on 14,000 street images) was able to indicate the correct images as effectively as a human.
Even when it made a mistake, the next attempt with a new puzzle was successful, and multiple attempts are allowed here. Moreover, the AI’s success rate did not decrease even after additional CAPTCHA security measures such as mouse movement analysis or the "user's" browser history were activated. The AI effectively mimicked a human enough to trick the system, raising serious security concerns.
The developed research is a significant signal for administrators responsible for security in online services. Although we are talking somewhat about academic considerations, it does not change the fact that the technology can be implemented in practice. Therefore, it is worth taking a closer look at website security and considering such modifications to the systems so that AI cannot overcome them. However, given the pace of AI development, this may prove to be quite a challenge.
