"Can You Find Vulnerabilities for Me?"... AI Agents Team Up to Find and Infiltrate Internal Networks on Their Own

AhnLab Tests an Autonomous AI Penetration Framework
Autonomously Executes Programs by Exploiting Web Server Vulnerabilities

A simulation has confirmed the threat posed by autonomous cyberattacks in which artificial intelligence (AI) independently carries out tasks ranging from vulnerability scanning and breaching external servers to moving through internal networks.

Yonhap News reported on October 18, citing cybersecurity industry sources, that the AhnLab Security Intelligence Center (ASEC) had released the results of a simulated penetration test to assess the real-world threat level of ARTEX, an open-source AI autonomous penetration framework.

ARTEX is a framework that uses large language models (LLMs) to coordinate multiple AI agents, each assigned a different role, in carrying out penetration tasks. Once an attack target and permitted scope are set, it sequentially handles tasks such as asset reconnaissance, vulnerability scanning, attack tool selection and verification of execution results. Rather than requiring the user to give instructions at every step, it uses information gathered from earlier tasks to determine what to do next.

In the experiments, ARTEX scanned for vulnerabilities, independently determined attack steps and executed programs. In a separate environment, it also succeeded in running remote commands on four internal hosts.

The photo is not directly related to the content of this article. Pixabay

AhnLab tested ARTEX in two simulated environments. In the first experiment, ARTEX investigated a web server with a vulnerability in its file upload function, installed a web shell and ran a program in a Windows environment.

In the second experiment, it attempted initial access by exploiting a vulnerability (CVE-2023-46604) in Apache ActiveMQ 5.15.15, a message broker software. It then investigated internal assets and accounts and ran remote commands on four internal hosts. The two experiments were conducted to determine whether AI could continue its tasks without receiving instructions for each stage of an attack.

AhnLab also investigated internet-exposed infrastructure associated with ARTEX. As of October 14, more than 900 public IP addresses were identified as belonging to systems running the framework. Materials related to attack activity, including web shells, internal network reconnaissance, credential collection and remote-control malware, were also found in some associated working directories.

The experiments show that AI-based penetration tools can link multiple stages of an attack, from vulnerability scanning and initial access to moving through internal networks. Unlike traditional methods, in which attackers had to carry out each task themselves, autonomous AI tools can choose their next actions based on execution results and continue the attack process.

AhnLab said, "AI agents are automating the entire process without interruption, from service discovery to privilege escalation and internal network penetration." It stressed that companies and institutions should inspect internet-exposed services and management functions and promptly address known vulnerabilities and security misconfigurations. It added that analyzing file and process execution records, network connections, web access logs and authentication logs together is important for detecting attack flows that extend from initial access to movement through internal networks.

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