AI/TLDR

Genians · 2026-08-10 · major

Kimsuky ran local LLMs on its own C2 servers — Genians details Operation GitPower

Genians Security Center found Ollama, GPT4All and Msty installed on servers run by Kimsuky, a North Korean espionage group. The same campaign, named Operation GitPower, uses AI-written decoy PDFs as phishing lures.

Genians threat intelligence report cover for the Kimsuky AI and local LLM analysis

Genians found three offline LLM runners installed on Kimsuky's attack servers, next to AI-written phishing lures.

Quick facts

Reported byGenians Security Center
PublishedAugust 10, 2026
CampaignOperation GitPower
GroupKimsuky, North Korea's Reconnaissance General Bureau
Local model runners foundOllama, GPT4All, Msty
MalwareEncrypted AsyncRAT delivered over Git-based C2
TargetsDiplomatic missions, military, security, virtual asset sectors

What is it?

Operation GitPower is the name Genians Security Center gives to an ongoing Kimsuky campaign in which the North Korean group set up local LLM environments on its own infrastructure. Genians found Ollama, GPT4All and Msty installed there, along with a configured retrieval-augmented generation database and copies of OpenAI's Whisper speech-to-text files.

How does it work?

The intrusion chain stays deliberately plain. A ZIP archive carries an LNK shortcut dressed up with a document icon and a business-sounding filename; opening it runs an obfuscated PowerShell loader. Git repositories then serve as the command-and-control channel and hand over encrypted AsyncRAT payloads disguised as image files. The AI work sits beside that chain rather than inside it.

Why does it matter?

Kimsuky used to recycle documents stolen in earlier breaches as decoys, which gave defenders a known file to fingerprint. Genians reports that since 2026 the group writes fresh ones with generative AI instead, so content-based phishing filters have far less to match against. That pushes detection toward behavior — what the LNK file and PowerShell actually do on the machine.

Who is it for?

security teams and threat hunters

Frequently asked questions

Is Kimsuky training its own AI models?
No. Genians assessed that Kimsuky is in a phase of gathering technology and capability rather than building models of its own. The group installed off-the-shelf local model runners and experimented with retrieval-augmented generation, the Cursor editor and OpenAI's Whisper speech-to-text files, which came with a guide for pulling text out of audio.
Has the offline AI setup been used in real attacks yet?
Genians has not confirmed the local LLM stack being run against victims. What it did observe in live attacks is the output side: decoy PDFs on virtual assets, financial investment and game development that Genians assesses were written with generative AI and sent as spear-phishing lures since 2026.
How does Operation GitPower relate to older Kimsuky activity?
Operation GitPower is not a new standalone campaign. Genians describes it as a continuation of years of Kimsuky operations that shares key traits with the FlowerPower campaign disclosed in 2023, including the PowerShell execution framework and the heavy abuse of Git repositories, now wired into a Git-based command-and-control structure.
What should defenders look for if the lure documents look clean?
Genians recommends behavior-based detection over document inspection. Its list covers abnormally long execution arguments inside LNK files, custom Base64 decoding, hidden PowerShell runs, scheduled tasks, calls to the GitHub Raw Contents API, personal access tokens unrelated to business work, and encrypted .NET payloads hidden behind image file extensions.

Sources · 3 outlets

Tags

  • security
  • threat-intelligence
  • kimsuky
  • north-korea
  • operation-gitpower
  • local-llm
  • ollama
  • gpt4all
  • rag
  • phishing
  • asyncrat
  • powershell
  • genians

← All releases · Learn AI