“Attack only designated targets”… AI penetrates internal network with one command line

“Infiltrate the attack target. Attack only the designated target.”

It was found that an artificial intelligence (AI) autonomous infiltration tool found vulnerabilities on its own through this short natural language command that even a person without professional hacking knowledge could write, penetrated an external web server, serially infiltrated up to four internal network hosts, and executed remote commands.

According to the security industry on the 18th, the AhnLab Security Intelligence Center released the results of a mock infiltration experiment that verified the actual threat level of ARTEX, an open source AI autonomous infiltration framework, through a recently released report.

Artex is an autonomous infiltration tool that operates based on the macro language model multi-agent.

Even without humans giving detailed commands, the agents responsible for setting goals, establishing plans, and executing tools are organically connected and autonomously repeat ‘reconnaissance → vulnerability search → attack route establishment → tool execution and verification’.

As a result of AhnLab testing Artex's penetration limits in a separate simulated environment, the AI ​​was able to find the entry point on its own and complete the attack even though no specific vulnerability or service name was provided.

In a web server environment, the user discovered a file upload loophole within the page, uploaded a web shell, and executed an arbitrary program.

In another corporate software exposure environment, after initial access to the perimeter server through its own software vulnerability, internal network assets and accounts were investigated and remote commands were executed on four internal network hosts without human intervention.

It was also revealed in detail that Artex was being abused as an actual attack infrastructure beyond simple concept verification.

As a result of AhnLab tracking the characteristics of the Artex web console, it was found that the number of related Internet Protocols (IPs) exceeded 900 as of the 14th.

In the previous analysis of all 600 IPs, the United States was concentrated at 53.7%, China at 19.8%, Hong Kong at 11.2%, and Singapore at 5.5%.

In particular, two open directories that appear to have been used by attackers to explore and infiltrate internal networks and relay remote desktop connections were discovered on some host servers where Artex was installed.

AhnLab secured a total of 51 files from these routes, including scripts and malware that can take over the internal network across Windows and Linux.

The data included scan results of 85 active systems in the Kubernetes environment and a script to extract account information from system memory dumps.

AhnLab warned that existing fragmented security detection methods have reached their limits as autonomous AI hacking has become a reality.

As a result of actual experiments, Artex left different web browser identification information values ​​depending on the function and calling tool, or erased traces by using a network method other than general web communication.

This means that it is difficult to detect and block AI-based attacks using only a specific network string or tool name.

AhnLab emphasized that "AI agents are seamlessly automating the entire process, from service discovery to authority theft and internal network infiltration," and added, "In addition to basic measures to reduce unused services exposed to the outside and quickly address vulnerabilities, it is essential to build an expanded detection and response system that organically links terminal detection and response solutions, networks, and authentication logs to identify the entire attack flow early."