Soongsil University Develops 'Ironclad' Server Security Technology That Tricks AI Hackers

Disrupting AI Attacks to Better Protect Real Servers

Researchers at Soongsil University have developed a security technology that lures AI hacking programs to fake servers to protect real servers. In an experiment, there were no server breaches after the defense technology was applied, compared with 35 breaches before its application.

From left: Professor Choi Daeseon of Soongsil University, master's student Noh Seungdeok, undergraduate student Ryu Seokjun, and research professor Park Sohee. Soongsil University

Soongsil University said on November 13 that a research team led by Professor Choi Daeseon of its AI Safety Research Center had developed a “deception defense” technology that tricks AI hacking tools—which autonomously identify and exploit security vulnerabilities—and blocks them from accessing real servers.

The research team tested the technology against ARTEX, an autonomous penetration-testing AI that Microsoft said had been used in state-sponsored hacking. In the experiment, 35 of 60 vulnerable web servers were breached when no defense technology was in place. After the new technology was applied, there were zero breaches of real servers.

Rather than blocking attacks, the technology developed by the research team deceives attackers. It works by analyzing a web service’s source code and pre-positioning fake access paths, pages, and passwords. When an attacking AI accesses one of these decoys, the system identifies the access as an attack and routes subsequent requests to a separate cloned server configured to resemble the real one. The attacking AI believes it has successfully penetrated the system, but in reality, it operates in an isolated fake environment.

The research team named the technology “Gingerbread House,” inspired by the fairy tale “Hansel and Gretel.” Unlike conventional methods of luring attackers to fake servers, the technology is designed to prevent anyone identified as an attacker from accessing the real server. Installing the decoys in advance took about $2 and 7–16 minutes per service.

The experiment also found that attacking AIs mistook fake passwords for real ones. In 30 of 110 valid trials, the AIs reported fake passwords as genuine; in 20 of those cases, the decoys alone fooled them without any separate penetration attempt. In 117 trials involving ordinary web crawlers, there were no instances of normal access being mistakenly identified as an attack. Breaches of real servers by attack agents using general-purpose AI models also fell from 26 to one.

The research team noted, however, that it conducted the experiment only once per task and did not test cases in which attackers anticipated the deception and changed their attack methods. It plans to assess the technology’s applicability using a range of attack tools and in real-world service environments.

“In an era when attacks move faster than patches, we need defenses that remain effective even after a breach,” Professor Choi said. “This study confirmed that deception defense is effective against AI hackers.”

This content was produced with the assistance of AI translation services.