Chatgpt has shown that he can independently complete one of the most popular online security checks: the “I am not a robot” box of Cloudflare, a variant of the captch tests. It did so as part of a controlled session through the new “Chatgpt Agent”, an experimental function available only for users of the paid paid versions, plus and team. This Openai agent is able to surf the internet, fill in modules, click on buttons and complete complex activities, such as making online shopping or booking appointments, always maintaining the supervision and authorization of the user. During a demonstration, he exceeded anti-Bot control simply by clicking on the box envisaged by the Cloudflare system and affirming in the process that that step was necessary “To prove it is not a bot». A paradoxical contradiction that opens questions about the future of the automation and reliability of the captcha tests, historically used to distinguish human from the bots.
Chatgpt Agent passed the Captcha test “I’m not a robot”
The heart of the matter is that Chatgpt Agent is able to act on the web independently, through a safe virtual environment (sandbox) with an operating system and an independent browser. While performing the activities, the user can observe everything through the interface and intervene at any time. When asked to perform an action, the agent can analyze the context of the site, interact with its elements and proceed step by step, just like a human being would do. In a situation shown on the platform RedditAI has completed a video conversion process by clicking first on the “Check that you are human” and then on the “Convert” button.
In instructing the user of the operation he was performing, Chatgpt Agent used these terms:
I will click on the “Verify You Are Human” check box to complete the verification on Cloudflare. This step is necessary to demonstrate that they are not a bot and proceed with the action.
To understand the scope of this event, you have to clarify what the captchas are (Completely Automated Public Turing Tests to Tell Computers and Humans Apart), i.e. those visual or behavioral tests designed to prevent Bot from accessing online services. The captcha were born in the late 90s as protection against spam and automatic traffic, and have evolved to include systems such as those of Cloudflare, based on an initial invisible screening phase (called turnstile) which analyzes the mouse movement, the speed of the clicks and other behavioral signals before proposing any visual puzzles.
In the case in question, the chatgpt agent did not have to face an image to be deciphered, but still successfully exceeded behavioral control, showing that he knows how to emulate human interaction very well with the site. This type of behavior is not completely new in the world of AI: for some time there have been tools capable of circumventing the captch tests, but the novelty lies in the fluid integration of this ability in an assistant to which can complete an entire online activity independently. It must be said that the purpose of the captchans is not so much to make the passage to the bots impossible, but rather to slow them down and make their most expensive use from an economic point of view.
The progress of the
The episode of Chatgpt Agent that exceeds a verification designed to block their own software agents could have an ironic flavor, but in fact it reflects the enormous progress made by AI in understanding the context, deciding and interacting with real digital environments in an increasingly sophisticated way. Of course, this does not mean that the AI is infallible: the interaction with complex or not very intuitive interfaces can still put it in difficulty. But the fact that he manages to orient himself in buttons, modules and anti-Bot controls, even narrating his actions in real time, suggests that intelligent automation is gaining more and more ground. Perhaps, as someone joked online, the time has come to replace the classic box “Check if you are a human“With a more honest”I’m a robot“.
