North Korean cyber-espionage group Kimsuky appears to be moving beyond occasional use of public AI services by assembling a local artificial intelligence environment that could eventually support phishing, data analysis and malware development.
South Korean cybersecurity firm Genians identified the setup after months of monitoring infrastructure linked to Kimsuky, a hacking unit subordinate to North Korea's Reconnaissance General Bureau. Its investigation found multiple tools capable of running AI models locally, alongside document-retrieval software, AI development libraries, speech-to-text components and an AI-assisted coding environment.
Genians found no evidence that Kimsuky had trained its own AI model. Instead, the activity indicates that the group is experimenting with existing technologies and assembling the components needed to incorporate AI into its established espionage workflow.
Among the strongest evidence were Ollama, GPT4All and Msty, applications that can run or interact with language models locally. Genians said the tools appeared to have been configured or used rather than merely downloaded. Ollama had generated keys associated with its initial setup, while GPT4All contained a configured "localdocs_v3.db" database used by its LocalDocs retrieval-augmented generation (RAG) feature.
RAG allows an AI model to retrieve information from a private document collection while generating responses. The database indicates an attempt to connect documents available to the operator with an AI system, although Genians could not establish that the documents were stolen.
Researchers also recovered an operator request asking for a dataset to be examined for cryptocurrency wallet information, Gmail credentials and website-registration history. The request demanded detailed analysis, but investigators could not confirm that it had actually been submitted to an AI service.
The infrastructure contained further evidence of AI development. Genians found LLamaSharp, Microsoft's Semantic Kernel and Microsoft.Agents.AI, frameworks that can help developers integrate AI capabilities into C# and .NET applications. OpenAI Whisper files and documentation for extracting text from audio were also recovered, alongside traces of Cursor, an AI-powered coding editor.
None of these tools is inherently malicious. Their relevance comes from their deliberate assembly on infrastructure associated with a nation-state espionage group. Together, they could provide building blocks for automating software development, document processing, transcription and information analysis.
The activity is linked by Genians to Operation GitPower, a Kimsuky campaign that has used GitHub repositories as command-and-control channels in an LNK-to-PowerShell infection chain and distributed encrypted AsyncRAT payloads disguised as image files. Fortinet has separately documented GitHub-based command-and-control activity targeting South Korean users, supporting the wider technique family, although it does not independently verify Genians' newly discovered AI artifacts.
AI could also make Kimsuky's phishing campaigns harder to identify. Generative systems can reduce linguistic weaknesses such as poor grammar, awkward translations and formatting errors that defenders have historically used as phishing indicators.
Kimsuky has already demonstrated interest in generative AI. In 2025, Genians reported that the group used ChatGPT to create a fake South Korean military identification card for a spear-phishing campaign.
The latest discovery suggests a further step toward keeping AI capabilities within infrastructure controlled by the attackers. Local models can process information without relying on public AI services, potentially giving operators greater control over sensitive material and reducing dependence on external platforms.
Kimsuky has historically targeted governments, researchers, think tanks, academics and other organisations for intelligence collection. The U.S. Treasury sanctioned the group in 2023 and described it as subordinate to North Korea's Reconnaissance General Bureau.
However, the findings should not be overstated. Genians has not demonstrated that the local AI stack has been deployed against victims, trained a proprietary model or autonomously developed malware. No GitPower victim count has been disclosed, and Reuters reported that the findings could not be independently verified.
For defenders, Genians recommends correlating LNK execution, PowerShell activity, hidden scheduled tasks, GitHub communications and subsequent payload execution rather than relying primarily on the quality of a phishing email.
The immediate development is therefore less about autonomous AI hacking and more about preparation. Kimsuky appears to be assembling the infrastructure that could allow AI to become an integrated layer across its existing phishing, malware and intelligence-collection operations.
Read the original article: