AI RiskThu, 22 Oct 2026
④ ARTEX: When attackers use agents to probe services and move inside a network
Evidence from attacks on financial organizations and a separate lab exercise shows how agent tooling can accelerate reconnaissance and execution.
Article
CrowdStrike reported on October 7 that attacks against South Korean financial organizations in late September and early October led to data exfiltration. It obtained ARTEX configuration files, Claude Code activity and memory files from an attacker-controlled directory. The main model backend for that ARTEX instance was DeepSeek v4.1-flash. Separate Claude Code sessions used GLM-5.3 and Grok 4.6.
On October 12, AhnLab reported finding web shell, internal reconnaissance, credential collection and remote control artifacts, with some execution traces. Its October 14 controlled exercise separately showed how ARTEX tested file-upload weaknesses or known ActiveMQ vulnerabilities and used tool output to choose subsequent actions.
- 01
Probe reachable services
The controlled exercise mapped web functions and backend services.
- 02
Exploit an entry point
The team separately tested file upload and known ActiveMQ vulnerabilities.
- 03
Move inward
The agent used tool results to probe the internal network and run additional commands in the lab.
What to do: from the perimeter to internal systems
- 01
Find exposed services
Inventory public, internal, outsourced and standby services. Patch reachable components first.
See what to doResources: 1
TWISA member solutions
Find exposed services
Inventory public, internal, outsourced and standby services. Patch reachable components first.
- 02
Test the paths
With authorization, test web uploads, APIs, messaging services and segmentation independently.
See what to doResources: 1
TWISA member solutions
Test the paths
With authorization, test web uploads, APIs, messaging services and segmentation independently.
- 03
Trace process activity
Correlate web service child processes, files and outbound connections to identify unexpected execution.
See what to doResources: 1
TWISA member solutions
Trace process activity
Correlate web service child processes, files and outbound connections to identify unexpected execution.
- 04
Investigate and recover
Preserve host and network records and review credentials, lateral movement and potential data access before recovery.
See what to doResources: 1
TWISA member solutions
Investigate and recover
Preserve host and network records and review credentials, lateral movement and potential data access before recovery.
Public services, hosts, operational credentials and internal data
Allow authorized traffic and apply controls at each layer.
- 01
Find exposed services
Public services, hosts, operational credentials and internal data
Validate the controlInventory public, internal, outsourced and standby services. Patch reachable components first.
- 02
Test the paths
Public services, hosts, operational credentials and internal data
Validate the controlWith authorization, test web uploads, APIs, messaging services and segmentation independently.
- 03
Trace process activity
Public services, hosts, operational credentials and internal data
Validate the controlCorrelate web service child processes, files and outbound connections to identify unexpected execution.
Feed logs and traffic data from each layer into monitoring and response.
- Investigate and recoverPreserve host and network records and review credentials, lateral movement and potential data access before recovery.
TWISA member solutions
Find exposed services
Inventory public, internal, outsourced and standby services. Patch reachable components first.
Test the paths
With authorization, test web uploads, APIs, messaging services and segmentation independently.
Trace process activity
Correlate web service child processes, files and outbound connections to identify unexpected execution.
Investigate and recover
Preserve host and network records and review credentials, lateral movement and potential data access before recovery.
Continue the discussion at CIX
A future CIX discussion can use one reachable service as a starting point for threat modeling. Bring a map of what an attacker could reach next and compare independent controls across web services, hosts and network segments.
Sources and further reading
- CrowdStrike investigation
- AhnLab analysis of observed artifacts
- AhnLab controlled experiment
- Read the full article in Traditional Chinese
Column updated: 2026/10/25.