Scientists Made an AI Model That Beats CAPTCHAs Every Single Time

Sep 30, 2024

Matt Growcoot

Two CAPTCHA verification challenges are shown. The left one asks to "Select all squares with traffic lights," and the right one asks to "Select all images with crosswalks." Both feature a grid of images where the user must identify specific elements, like traffic lights and crosswalks, respectively.

Researchers have made an AI bot capable of beating image-based CAPTCHAs 100 percent of the time, according to a new scientific paper.

Internet users will be familiar with image-based CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart) that show a grid of images requiring the user to select the panels with objects such as motorcycles or traffic lights contained within. CAPTCHAs are a security measure that helps to prevent spam and bot attacks by veryfing the user is human.

But thanks to AI technology developments, PhD student Andreas Plesmer and his colleagues at ETH Zurich University in Switzerland were able to build a model capable of beating Google’s ReCAPTCHA v2 system every time.

Ars Technica reports that Plesner used a fine-tuned version of the open-source YOLO (You Only Look Once) object-recognition model which was previously used by a clothing line to confuse AI cameras.

To beat Google’s reCAPTCHA v2, scientists trained the model on 14,000 labeled images so it could learn the objects likely to show up on the CAPTCHA. The model can identify pictures of a motorcycle 69 percent of the time but for fire hydrants, it was 100 percent of the time. Regardless, the model was able to pass the CAPTCHA test every time even when given multiple challenges.

Internauts may have noticed that image-based CAPTCHAs are appearing less and less, that’s because Google began phasing out the picture CAPTCHA system years ago in favor of reCAPTCHA v3; an invisible security measure that analyzes user interactions to decide if it’s human rather than giving a specific challenge. But Ars Technica notes that reCAPTCHA v2 is still used by millions of websites and reCAPTCHA v3 still uses v2 as a fallback option when it suspects the user is a bot.

Previous studies attempting to beat CAPTCHAs with image-recognition models have had a success rate of around 70 percent. Plesmer and his colleagues say the rise to 100 percent “shows that we are now officially in the age beyond CAPTCHAs.”

“In some sense, a good CAPTCHA marks the exact boundary between the most intelligent machine and the least intelligent human,” the authors’ write. “As machine learning models close in on human capabilities, finding good CAPTCHAs has become more difficult.”