Kimsuky Prepares an Offline AI Stack to Enhance Phishing, Malware, and Data Collection

Activity Attributed to the North Korean Group

Genians attributes the preparation of infrastructure for running artificial intelligence tools offline to Kimsuky, a North Korean unit subordinate to the Reconnaissance General Bureau.

The activity was identified on August 10, 2026, following months of monitoring and log analysis. The evidence points to a research, assembly, and testing phase—not the training of a proprietary model.

The apparent goal is to apply AI across multiple operational stages: malware development, phishing campaign creation, and automated data analysis. There is no evidence that the stack has already been used against a victim.

Tools Identified on the Monitored Systems

The observed components include several software products for running models locally or integrating them into applications:

• Ollama, with keys generated during first launch, suggesting that it was actually run or configured;
• GPT4All, with the localdocs_v3.db database associated with the LocalDocs feature and retrieval-augmented generation (RAG);
• Msty, a platform for running language models locally;
• LLaMaSharp, Microsoft Semantic Kernel, and Microsoft.Agents.AI, libraries for AI applications in C#/.NET environments;
• OpenAI Whisper, through files and instructions for transcribing audio content;
• Cursor, a development editor with AI-assisted features.

The exact product versions have not been disclosed. The RAG database shows an attempt to connect documents available to the operator to an AI system, but does not prove that those documents were stolen.

In one activity recovered from the logs, the operator requested a dataset check for information on wallets, Gmail credentials, and website registration history. It was not possible to confirm whether the request was actually sent to an AI service.

Link to Operation GitPower

The activity is part of Operation GitPower, a campaign that uses GitHub repositories as command-and-control channels. The observed infection chain follows the LNK → PowerShell pattern and delivers encrypted AsyncRAT payloads disguised as image files.

In April, Fortinet separately documented the use of the GitHub-C2 pattern against South Korean users. That report confirms the general technique but does not establish a connection to the new local AI artifacts.

Genians attributes the operation to Kimsuky based on overlaps with previous campaigns, infrastructure elements, and North Korean terminology found in the logs. In 2023, the U.S. Treasury Department sanctioned the group, describing it as an organization primarily focused on intelligence collection.

In 2025, Genians also linked Kimsuky to a spear-phishing campaign that used ChatGPT-generated images depicting South Korean military employee identification cards.

Why Offline Execution Changes Detection

A local infrastructure allows documents and data to be processed without sending them to public chatbots. It can also make malicious messages more convincing by reducing unnatural translations, typos, and formatting issues.

Text quality alone is therefore no longer sufficient to identify a phishing attempt. The strategic risk is significant, but full operational automation has not yet been demonstrated, and no victim count for GitPower has been disclosed.

What Security Teams Should Look For

Detection efforts should correlate the following events within the same sequence:

• execution of LNK files and PowerShell activity;
• creation or activation of hidden scheduled tasks;
• traffic to GitHub repositories;
• downloads of encrypted payloads disguised as images;
• subsequent execution of the payloads.

Teams should also monitor the installation, launch, and configuration of Ollama, GPT4All, Msty, Cursor, and Whisper, as well as AI libraries for C#/.NET.

Hunting should include local models, keys generated during first launch, RAG databases such as localdocs_v3.db, and indexed documents. Triage should prioritize process chains, persistence, communications, and post-compromise activity rather than relying solely on the linguistic quality of phishing lures.