AI Autonomous Hacking Becomes Reality with Simple Commands

A recent report from AhnLab Security Intelligence Center revealed that an artificial intelligence (AI) autonomous penetration tool can identify vulnerabilities and execute remote commands, including breaching external web servers and infiltrating internal networks, all with a simple command.

The report, released on October 18, details the results of a penetration test that validated the actual threat level of ARTEX, an open-source AI autonomous penetration framework.

ARTEX operates based on a multi-agent large language model, allowing agents to autonomously set targets, plan, and execute tools without detailed human commands. This enables repeated autonomous reconnaissance, vulnerability scanning, attack path establishment, tool execution, and verification.

AhnLab tested ARTEX's penetration limits in a controlled environment. Although specific vulnerabilities or service names were not disclosed, the AI successfully identified entry points and completed attacks. In a web server environment, it autonomously discovered a file upload vulnerability, uploaded a web shell, and executed arbitrary programs.

In an enterprise software exposure scenario, the AI gained initial access through a software vulnerability in a perimeter server, then investigated internal assets and accounts, executing remote commands on four internal network hosts without human intervention.

Evidence suggests that ARTEX is being misused as an actual attack infrastructure, going beyond mere proof of concept.

AhnLab tracked the characteristics of the ARTEX web console, finding that as of October 14, over 900 Internet Protocol (IP) addresses were associated with it. An earlier analysis of 600 IPs revealed that 53.7% were from the United States, followed by 19.8% from China, 11.2% from Hong Kong, and 5.5% from Singapore.

Some host servers with ARTEX installed showed signs of attackers using open directories for internal network exploration, penetration, and remote desktop connection relays.

AhnLab secured 51 files, including scripts and malware capable of taking control of internal networks, which traversed both Windows and Linux systems. These files contained results from scanning 85 active systems within a Kubernetes environment and scripts that exported account information from system memory dumps.

As AI autonomous hacking becomes a reality, AhnLab warns that existing fragmented security detection methods have reached their limits. ARTEX can obscure traces by passing different web browser identification values or utilizing alternative network methods instead of standard web communications. This indicates that detecting and blocking AI-based attacks cannot rely solely on specific network strings or tool names.

AhnLab stated, "AI agents are automating the entire process from service exploration to privilege escalation and internal network infiltration without interruption. It is essential to reduce unused services exposed to the outside, quickly address vulnerabilities, and establish an extended detection and response system that integrates endpoint detection and response solutions with network and authentication logs to identify the entire attack flow early."