Image to help understand the article
SEOUL, South Korea — An open-source artificial intelligence tool breached an externally facing web server and moved through a simulated corporate network to execute remote commands on four internal computers after receiving only a short, plain-language instruction, according to research by South Korean cybersecurity company AhnLab.
AhnLab’s Security Emergency Response Center, known as ASEC, tested the autonomous penetration framework ARTEX in a separate, controlled environment. The report, recently released by the company, became known in the cybersecurity industry on the 18th.
The prompt told ARTEX only to infiltrate the target and limit its attack to designated systems. Researchers did not identify a vulnerability or explain what services were running. The tool found its own entry points and completed the attacks without human intervention, AhnLab said.
From Initial Access to an Internal Network
In one test, ARTEX identified a file-upload weakness on a website, installed a web shell — malicious code that allows remote control of a server — and ran an arbitrary program on the machine.
In another environment containing enterprise software, ARTEX exploited a known Apache ActiveMQ vulnerability, tracked as CVE-2023-46604, to gain access to a perimeter server. It then scanned the internal network, investigated systems and accounts, and moved laterally across the network, ultimately executing remote commands on four hosts.
ARTEX combines a large language model with three specialized AI agents. A Goals agent organizes the objective, a Planner chooses each next step based on previous results, and a Worker runs tools and attempts exploits. The framework continuously links an asset graph containing information about targets with an exploration graph that tracks objectives and discovered vulnerabilities.
Once a user defines the target and scope, the agents can repeatedly perform reconnaissance, search for vulnerabilities, develop attack paths, run tools and verify results without requiring step-by-step instructions.
Infrastructure Concentrated in the United States and Asia
AhnLab also tracked the characteristics and hosting addresses of ARTEX web consoles exposed to the internet. More than 900 related internet protocol addresses had been identified as of the 14th.
In an earlier review of 600 public IP addresses that could be examined in detail, 53.7% were in the United States. China accounted for 19.8%, Hong Kong for 11.2% and Singapore for 5.5%. Together, those four locations represented about 90.2% of the addresses.
The figures indicate where the infrastructure was hosted, not necessarily where its operators were located.
Evidence of Real-World Attack Preparations
Researchers found two openly accessible directories on some ARTEX host servers. AhnLab said the locations appeared to support internal network discovery, intrusion, pivoting between systems, relaying Remote Desktop Protocol connections and deploying additional malicious software.
AhnLab recovered 51 files, including scripts and malware that could help compromise both Windows and Linux networks. They included a script for issuing commands through a web shell, a file for relaying remote desktop connections and a script designed to extract credentials from heap dumps.
The material also included scan results covering 85 active systems in a Kubernetes environment and a script for retrieving account information from system memory dumps. AhnLab said the findings suggest ARTEX has moved beyond a proof of concept and is being used to prepare and carry out actual attacks.
Traditional Warning Signs May Not Be Enough
During testing, ARTEX generated different browser identification strings, known as User-Agent values, depending on the function and tool it used. It could also conceal activity through alternative network methods, while an operator could change or remove those identifiers.
AhnLab warned that security teams cannot reliably detect such attacks by watching for a single header value, network string or tool name. The company said autonomous AI hacking is already a reality and that defenses focused on isolated indicators have reached their limits.
AhnLab recommended reducing unused services exposed to the internet, quickly patching vulnerabilities and connecting endpoint detection, network monitoring and authentication logs through an extended detection and response system. That broader view, it said, is necessary to identify the full course of an AI-driven attack before it spreads through an internal network.
Comments
Post a Comment