Genians finds Kimsuky assembling offline AI tools for espionage
South Korean security firm Genians has identified an offline artificial-intelligence toolset on infrastructure it attributes to Kimsuky, a North Korean espionage group linked to the Reconnaissance General Bureau. The environment included Ollama, GPT4All and Msty, alongside developer components and transcription tooling that could support AI-assisted analysis, phishing and malware development.
Genians said the finding followed months of infrastructure tracking and log analysis. It found no evidence that Kimsuky trained a model of its own. Instead, the firm characterises the activity as research and knowledge acquisition: assembling, configuring and testing available software with an apparent aim of extending AI across the group’s operational workflow.
Evidence points to local models and document retrieval
The distinction between downloaded software and an operational experiment is important. Genians reported that Ollama generated the keys created on its first launch, while GPT4All contained a configured localdocs_v3.db database. That database supports GPT4All’s LocalDocs retrieval-augmented generation feature, which enables a model to answer questions using a private document collection.
The database shows an attempt to connect documents held by the operator to an AI system, but it does not establish that the documents were stolen. Researchers also recovered an operator request to inspect a data set for wallet details, Gmail credentials and site-registration history. Genians could not confirm that the request itself had been submitted to an AI service.
Development components broaden the potential workflow
On the same infrastructure, Genians found LLaMaSharp, Microsoft Semantic Kernel and Microsoft.Agents.AI, libraries used to add AI capabilities to custom C# and .NET applications. The researchers also identified OpenAI Whisper speech-to-text files and guidance for extracting text from audio, as well as active traces of the Cursor AI coding editor.
These products are widely available, but their combination on infrastructure tied to an espionage operation is notable. The observed activity has not been shown running against a victim, and Genians disclosed no victim count for its associated campaign. Reuters said the local-AI findings could not be independently verified.
GitPower remains the operational context
Genians connects the infrastructure to Operation GitPower, which abuses GitHub repositories as command channels within an LNK-to-PowerShell infection chain and has delivered encrypted AsyncRAT payloads disguised as image files. Fortinet documented the broader GitHub command-and-control pattern in April, corroborating the surrounding technique family rather than the newly reported local-AI artifacts.
The development also follows AI-agent incident trends on AI-agent incident trends, while underscoring a more concrete defensive issue: a polished lure is becoming a weaker signal of malicious intent. Genians recommends correlating LNK execution, PowerShell activity, hidden scheduled tasks, GitHub traffic and subsequent payload behaviour.
For security teams, the practical implication is to prioritise endpoint and process telemetry around the full execution chain, rather than relying chiefly on spelling errors, formatting flaws or other traditional signs of a phishing message.