2026-10-31 18:52 UTC
----------------
🎯 Threat Intelligence
===================
Zscaler ThreatLabz published a technical walkthrough of the infrastructure behind an AI-assisted hacking campaign, building on Recorded Future's earlier report on TAG-217 activity against South Korean banks. The recovered material is heavy on operational detail, and the authors are explicit about what is confirmed versus inferred: model names in the configurations are reproduced as requested IDs and were not independently verified.
🔹 Two servers, one workflow
The work server stored project files, scripts, AI settings and a local ARTEX deployment. The jump server, 38.244.50.120, was set as the default route for connections to targets and also ran additional AI sessions. The operator delegated research tasks to Grok on that server and ran a separate AI session whose job was to review Grok's findings. Work was split between the two servers, with the operator providing oversight and direction.
Three custom scripts kept the environments consistent:
• tools/sync-all.sh: uses SSH and rsync to copy AI instructions, reusable task guides, saved notes and tools to the jump server's root, claude, grok and gemini accounts
• sync-memory.sh: copies saved AI memory files to locations on the jump server
• tools/sync-cc-users.sh: copies shared Claude Code instructions and supporting files to the gemini and glm accounts on the work server
The scripts also rewrote model references inside specialist agent instructions so shared instructions matched the destination account, and they included backups plus validation checks to avoid overwriting modified files.
🔹 Attack Chain Analysis
No malware infection chain is described in the source. What is documented is the operational chain:
1. Environment provisioning: work server holding project files, scripts and AI settings; jump server 38.244.50.120 as the default route to targets
2. AI orchestration: research delegated to Grok, a separate reviewer session, and sync scripts keeping instructions consistent across accounts
3. Task execution: customized ARTEX with a planner-and-worker structure, extended with web, network and Android tooling
4. Model routing: requests leave through OpenAI-compatible gateways such as xcai.pro, decoupling account names from the models served
🔹 Customized ARTEX deployment
ARTEX is an AI-assisted penetration-testing framework. Its planner-and-worker architecture lets an operator define a goal, which AI agents break into tasks, execute with tools and document for review. The recovered deployment extends the framework with additional tools for web, network and Android analysis.
The saved Docker Compose file defines an ARTEX container plus a PostgreSQL 16 database, mapping the interface to host port 8890 and PostgreSQL to 127.0.0.1:5435. The image is tagged artex-local:perf2-20261003. The Dockerfile adds nmap, sqlmap, nuclei, httpx, jadx, apktool and frida-tools, along with Python libraries for web processing and spreadsheet output, and replaces the stock ARTEX program with the operator's local version.
🔹 Model routing and the gateway problem
One detail deserves attention: account names did not always match the configured model, so the client application (such as Claude Code) must be distinguished from the model service receiving requests. The saved ARTEX environment configuration reads:
ARTEX_LLM_PROVIDER=openai
ARTEX_LLM_MODEL=deepseek-v4.1-flash
ARTEX_LLM_BASE_URL=https://xcai.pro/v1
Here openai denotes an OpenAI-compatible API format, the model field requests deepseek-v4.1-flash, and the base URL sends traffic to the xcai.pro gateway. The report is careful to stress that this confirms the requested model and endpoint, not the model the gateway actually served. Treat these identifiers as claims, not facts.
🔹 Detection notes
The source does not include detection rules, but the recovered artifacts give concrete hunting options:
• Outbound connections to 38.244.50.120
• API traffic toward xcai.pro
• Filenames such as tools/sync-all.sh, sync-memory.sh and tools/sync-cc-users.sh
• Docker deployments exposing port 8890 with image tag artex-local:perf2-20261003
🔹 Caveats
This is an infrastructure teardown, not a complete intrusion report. The article excerpt is truncated, so victim specifics and a timeline are missing here. Attribution context rests on Recorded Future's separate reporting on TAG-217 and South Korean banks.
References: Recorded Future's blog post on TAG-217 activity against South Korean banks; the Zscaler ThreatLabz analysis.
🔹 ThreatIntel #AI #TAG217 #ARTEX #PenetrationTesting
🔗 Source: https://threatlabz.zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
Replies (0)
No replies.