I saw artificial intelligence agents trying to penetrate my website

A few weeks Before, I saw a small team from artificial intelligence Agents spend nearly 10 minutes try to penetrate my new encoded website.

Artificial intelligence agents, which were developed by starting Runsybil, are working together to investigate my bad location to determine weak stains. The Orchestrator agent, called SYBIL, oversees many specialized agents, all of whom are supported by a set of customized language models and a ready -made application programming interface.

While traditional suburbs of the known known problems, Sybil is able to work at a higher level, using artificial intuition to know the weaknesses. It may be, for example, to work that the guest user has a distinct access – something that the regular scanner may miss – this is used to build an attack.

Ariel Herbert Foss, CEO and founder of Runsybil, says it is likely that the growing artificial intelligence models of artificial intelligence are offensive and defensive cybersecurity. Herbert Foss told me: “I would like to claim that we are certainly the threshold of a technical explosion in terms of capabilities that both bad and good actors can benefit from,” Herbert Foss told me. “Our mission is to build the next generation of offensive security tests just to help everyone keep up with.”

The website targeted by SYBIL was recently created using Claude Code to help me sort through new artificial intelligence papers. The site I call Arxiv Slurper It consists of the back interface server that reaches Arxiv –Where most artificial intelligence research is published– With some other resources, combing with paper summaries of words such as “Novel”, “First”, “Surprise”, as well as some technical terms that I care about. It is a continuous work, but I was impressed by the ease of packing together a possible useful thing, even if I had to fix some insects and hand composition problems.

There is a major problem with this type of encrypted location in elegance, however, it is difficult to know the types of security weaknesses that you may have provided. So when I talked to Herbert-Voss about Sybil, I decided to ask if he could test my new site to get weaknesses. Fortunately, only because my location is very essential, Sybil did not find any weaknesses.

Herbert-Voss says most of the weaknesses tend to be the result of more complicated functions such as models, additional components and encryption features. We saw the calculation of the agents who tried to investigate The fake e -commerce site With the known weakness of Herbert Fos. Sybil built a map of the application and how to access it, investigate weak spots by treating parameters and testing edge cases, then linked to chains together, testing hypotheses, and escalating until it breaks something meaningful. In this case, the site penetration methods were identified. Unlike a person, Herbert Foss says that Sibel runs thousands of these operations in parallel, and does not miss the details, and does not stop. He says: “The result is something that behaves like an experienced attacker, but it works accurately and its size,” he says.

“The pen that works with artificial intelligence materials is a promising trend that can have great benefits for defense systems,” says Logo Power, the computer world at Carnegie Mellon University (CMU), a specialist in computer security. Power recently written study With others from CMU and a researcher from AI Compane Anthropic, the promise to test the artificial intelligence penetration test. The researchers found that the most advanced commercial models were unable to perform network attacks, but they developed a system that defines high -level targets such as scanning a network or host injury, which enabled them to conduct hacking tests.