I Watched AI Agents Try to Hack My Vibe-Coded Website

Several weeks Before, I saw a small team Artificial intellect Agents spend about 10 minutes trying to hack my brand’s new vibbe-coded website.

AI agents developed by Startup Roonsabil worked together to search my weak site to identify poor spots. An orchestator agent, known as Sibil, supervises a custom language model and a number of other specialized agents powered by the off-shelf API combination.

Conventional weakness scanners search for specific known problems, to work at high levels using artificial intuition to determine the weaknesses of Cibil. It may be for example, a guest user is available for access – can miss any regular scanner – and it can use it to create an attack.

CEO of Ronsibil and Cafouder Ariel Herbert-Vos says that growing AI models can lead to both aggressive and defensive cyberquacy. Herbert-Vos told me, “I would argue that we certainly have a technology exploded in terms of skill that can take advantage of both bad and good actors.” “Our goal is to create the next generation’s aggressive protection test to help everyone continue.”

I recently created the website of Sibil using the Claud Code to help pick up the new AI research papers. Site, which I call Alexive Consisting of a back and server that accesses the AlexiveWhere most of the AI research is posted– including a few more resources, combining paper for words like “novels”, “first”, “surprise”, as well as some technical terms I am interested in it is going on, but I was fascinated how it was easy to see how easy I had to fix some bugs and configuration issues, but I was fascinated.

However, one of the main problems of this type of vibe-coded site is hard to know what kind of security weakness you can turn on. So when I talked to Herbert-Vos about Sibel, I decided to ask if it could test my new site for weakness. Thanks, and only my site is incredibly basic, Sibil didn’t find any weaknesses.

Herbert-Vos says that most vulnerabilities continue to result in more complex functioning such as forms, plug-ins and cryptographic properties. We saw the same agents trying to investigate A Dummy Ecommerce website With a familiar weakness owned by Herbert-Voss. Cibil created a map of the application and how it was accessed, the cases of the parameters and the cases of the test edges were searched for vulnerable spots, and then examined the search, the estimates and it enhanced anything meaningful. In this case, it has detected ways to hack the site. Unlike a man, Herbert-Vos says that Sibil runs in parallel to these thousands of processes, does not miss the details and does not stop. “The result is something that behaves like a ripe attacker but works with the accuracy and scale of the machine,” he said.

“The AI-powered pen test is a committed direction that can get significant benefits for defending systems,” said Luzo Bour, a computer scientist at the University of Carnegie Mellon (CMU) at AI and Computer Protection. Bauer recently colleague A study CMU with others and an ethnographic researcher at the AI company that explore the promise of AI intrusion test. Researchers have discovered that the most advanced commercial models cannot attack the network, but they have created a system that sets high-level objectives such as scanning a network or infection of a host, which enables them to test their infiltration.