Answering CAPTCHA puzzles is a pain, but until now, it made sense to use them as an (imperfect) defense against malicious bots. New research from ETH Zurich shows that artificial intelligence can now beat these puzzles every time. CAPTCHA is an acronym that stands for “Completely Automated Public Turing test to tell Computers and Humans Apart” and is used on a huge range of websites.
But the tool may have to be renamed based on how well an AI model created by Swiss researchers solves the security measure’s word and object identification puzzles.
The AI puzzle solver is built on a widely used image processing AI model called You Only Look Once (YOLO). Scientists tuned YOLO to compete with Google’s popular CAPTCHA reCAPTCHAv2 version. reCAPTCAv2 will quickly recognize you whenever you click on a car, bike, bridge, traffic light, etc. to prove you’re human.
But with 14,000 labeled road photos as training data and a little time, scientists were able to get YOLO to recognize objects as well as humans can. In fact, the AI is only as good as humans, since it didn’t solve every puzzle perfectly the first time. But it might remember that it has multiple chances unless it completely messes up the puzzle. YOLO was able to perform well enough that even if it made a mistake on one puzzle, it could make up for it and succeed on another CAPTCHA puzzle.
Narrowing the scope of objects that users need to identify (often just 13 categories, such as traffic lights, buses, and bicycles) allowed for easier integration across websites.
But this narrow focus on a narrow range of object types made it easy for the YOLO-based AI model to beat the system. The ETH Zurich team says that the system’s simplicity worked in its favor, allowing the AI to easily complete image-based challenges. Despite attempts to make the CAPTCHA more sophisticated by incorporating elements such as mouse movements and browser history (known as device fingerprinting), the AI’s success rate remained the same.
The rise of CAPTCHA-cracking AI
The fact that AI systems can now circumvent CAPTCHA systems with a perfect success rate is a wake-up call for the cybersecurity community. CAPTCHA systems are a critical component of web security, designed to prevent activities such as bots spamming, creating fake accounts, and launching distributed denial-of-service (DDoS) attacks. If these systems are compromised, websites can become more vulnerable to automated attacks and other malicious activity.
The success of the YOLO model in cracking CAPTCHA systems is not an isolated story. In recent years, AI models have become increasingly adept at tasks once thought to be only human capable of performing. Solving CAPTCHA puzzles is just the latest milestone in AI advancements that have redefined expectations around machine learning and automated systems.
Impact on everyday users
For the average person, CAPTCHA puzzles are encountered on a daily basis when logging into online accounts, submitting forms, shopping online, etc. The security of these interactions depends on the CAPTCHA’s ability to keep bots out. With these latest advances in AI, there is a real risk that CAPTCHA will no longer serve its intended purpose as an effective gatekeeper.
One immediate concern is that if CAPTCHA systems become outdated or are easily circumvented by bots, they could lead to an increase in automated activity such as spam and bad bot campaigns. For example, CAPTCHA systems are often used to prevent bots from creating thousands of fake accounts or automatically posting spam content across social media platforms. If bots can easily circumvent CAPTCHAs, this could lead to an increase in fraudulent activity across websites.
Moreover, defeating CAPTCHA techniques will force websites and service providers to seek stronger security mechanisms. Alternatives being considered include more advanced behavioral analytics techniques, such as tracking user interaction patterns, and biometric-based verification systems that rely on fingerprint or facial recognition.
Am I an AI?
Proving you’re not a robot isn’t as easy as it used to be, but don’t panic that it will be replaced anytime soon: this is simply proof that cybersecurity needs to take into account the rapidly evolving capabilities of AI models. CAPTCHAs may be phased out in favor of different puzzles to prove you’re human.
More thorough measures will be needed than simply choosing the right image. Security settings may need to monitor puzzle-solving behavior, such as typing and scrolling speed and accuracy. Or a combination of multiple tests and validations may be needed. In other words, cybersecurity will need to get stricter, but preferably without slowing down web browsing too much. If things get really strict, we’ll probably all shed a tear watching Mufasa die in The Lion King.
