RED Security SOC analysts linked the compromise of a large commercial company's domain to the use of the ARTEX AI agent. The practical challenge for defenders was not just the speed of the attack: the agent left behind hundreds of scripts and results, and several workstreams proceeded in parallel even after domain privileges were obtained.
In their analysis, the team describes an incident in which the attacker went from external reconnaissance to domain administrator-level access. At the same time, the authors call the conclusion about ARTEX a high-probability hypothesis, not an indisputably proven fact. According to them, it was the artifacts on the compromised systems that made it possible to reconstruct a significant part of the sequence of actions.
From a web application to the domain
The earliest trace discovered was a GET request to /.git/config on one of the company's public resources. After about four and a half hours of active reconnaissance, the attacker gained administrative access to two web application installations using default credentials. Through the standard file-upload feature, they placed PHP web shells and began executing commands in the containers.
The attacker then explored the internal environment. The application configurations contained credentials for LDAP, SSO, and other systems, and through OpenMetadata the attacker obtained SQL connection parameters for MSSQL servers. Linked server connections made it possible to execute xp_cmdshell on other servers; later, commands were executed on Windows, and access was escalated to SYSTEM.
A separate path ran through Active Directory. In SYSVOL, they found a group policy with a cpassword, from which they recovered an account password. An AD CS misconfiguration also opened the door to ESC1 exploitation: the attacker obtained domain administrator certificates and performed DCSync. The database dump was placed in an open directory of a compromised web application and downloaded.
According to RED Security, 7 hours 40 minutes elapsed from the start of reconnaissance to domain compromise, and 2 hours 14 minutes from the first successful POST request to the web shell. In parallel, the attacker also worked with other applications: changing roles and integrations, creating administrator accounts, and testing file reads through the API. Some of these scenarios continued for hours.
Scripts, parallel tasks, and traces of an agent
The main suspicions of an automated AI agent arose after analysis of the containers that served as the entry point. More than 600 files were found there: small scripts in different languages, their variants, intermediate data, and execution results. Based on timestamps, analysts determined that up to 13 scripts were created in some minutes, and up to 25 objects when auxiliary files were included.
The collection included Python, PHP, shell, PowerShell, and BAT files. PowerShell commands were often embedded in Python scripts or BAT files. Responses to requests and subsequent versions of checks were left alongside them — not just ready-made tools, but traces of iterations: an attempt, the result, an adjustment, and another run. Familiar tools such as Impacket and Certipy were also used, but were installed in the compromised container.
The output structure also caught their attention: results were grouped by targets and checks, with separators, headers, and success markers. A tidy format by itself does not prove AI involvement, but together with the pace of file creation and recurring structures, it became part of the body of evidence. Manual work leaves traces too; what was unusual here was their scale and the way they were coordinated.
Numeric labels in file names, accounts, and requests recurred across different systems. For example, the same label linked checks on several external applications, while another connected artifacts in a Linux container and on a Windows server. This end-to-end numbering helped match individual actions to specific workstreams.
Another distinctive feature was parallelism. At the same time, they were testing file-upload methods on different resources, document retrieval parameters, and API inputs. Even after obtaining a domain administrator account and performing DCSync, external exploitation attempts did not stop. To the analysts, this looked like work on several independent tasks, not the sequential execution of a single scenario.
Why the attribution was linked to ARTEX
They found no typical pattern of persistence through a separate tunnel or a familiar set of management tools in the containers. The main foothold remained the web shell in the initial application: commands were launched through it and some interactions with internal servers were proxied. On Windows systems, in particular, they found CGI and ASP.NET web shells; for one variant, a separate IIS pool and a local firewall rule for TCP/8888 were created.
An additional clue came from checking the external address in VirusTotal: the community had a label, ARTEX Found, and port 8787. ARTEX is described as a tool for attacks involving AI agents. The team then compared the observed file structures with the way intent tasks are organized in the agent's code: separate subtasks can run in parallel, record results, and finish independently.
This correspondence strengthens the ARTEX theory, but does not turn it into a definitive attribution. The report phrases the conclusion cautiously: the infrastructure was likely compromised using this agent. The material does not disclose the name of the company that was attacked.
What changes for the SOC
For the investigation, what mattered was not individual requests or a single web shell, but how they lined up over time: rapid script creation, repeated attempts after receiving responses, several active branches, and shared labels across different systems. These signs helped distinguish the described case from standard mass scanning, where a tool usually leaves a recognizable User-Agent. Here, internal requests went through compromised containers with User-Agent curl or python requests.
Analysts recorded about 1600 POST requests to the web shell over approximately three and a half hours. The large number of artifacts provided material for reconstruction, but also increased the workload: they had to determine where credentials came from, link successful checks to subsequent actions, and avoid losing secondary scenarios amid the fact that the domain had been taken over. Here, “more logs” does not automatically mean “easier to figure out” — sometimes it just means more context to keep track of.
RED Security reports that the team was able to reconstruct the main branches of the incident, the sources of credentials, the path to domain privileges, and the data compromise. The case authors emphasize that future work should take into account not only the speed of the attack, but also its high intensity and parallel task execution — both during monitoring and during digital forensics and report preparation.









