North Korean government hackers are running AI language models on local infrastructure to improve their cyberattack operations while avoiding cloud services that could expose their activities, according to a report published Monday by South Korean security firm Genians. The researchers observed the Kimsuky espionage group—operating under North Korea's Reconnaissance General Bureau—setting up local AI environments and collecting technology to integrate machine learning into malware creation, data analysis, and attack methods. The findings offer clear proof that the threat actors have moved past casual testing and are actively preparing to embed AI capabilities into real operations.
Genians identified Kimsuky establishing local language model environments using tools including Ollama, GPT4All, and Msty, while also experimenting with Cursor AI for code editing and testing retrieval-augmented generation (RAG) for searching documents stored locally. The group collected numerous libraries such as LLaMaSharp and Microsoft.Extensions.AI, plus packages including OpenAI and Azure.AI.OpenAI that enable integration of commercial AI services into custom applications. Researchers also uncovered logs showing the hackers used speech-to-text tools, including OpenAI's Whisper models. The group's recent phishing campaigns employ ZIP archives hiding malicious LNK files disguised as international event materials, research documents, or meeting invitations—when opened, these shortcuts execute embedded PowerShell loaders that harvest extensive system data including OS version, architecture, configuration, boot history, and running processes. In some instances, the attackers used AI to generate lure documents related to virtual assets and finance that employ natural language and polished formatting resembling legitimate business materials.
The researchers found that development components spanning "local AI execution → document retrieval (RAG) → automated agents → external AI integration" were collected together, which "strongly suggests that they were not gathered out of simple curiosity, but for the direct development of an AI-based tool designed for a specific purpose." Genians' investigation uncovered the spies using Git-based command-and-control infrastructure not just for attack operations but also for malware development, testing, stolen data management, and AI technology research. Multiple public GitHub repositories operated by the threat actors contained configuration files, PowerShell scripts, and various payloads used in follow-on attacks. The report notes that running AI models locally prevents conversation data from reaching external services, reducing exposure risk and making it especially appealing for state-sponsored hackers.
The attackers' choice to operate language models on their own controlled infrastructure rather than using cloud-based AI services allows them to experiment with malicious applications without leaving traces in commercial systems that intelligence agencies might monitor. By applying RAG technology to stolen files, the hackers can automatically and rapidly extract valuable information from large data volumes without manual review. Genians found no evidence that North Korea has begun training its own AI models—the focus remains on applying existing AI to malware development and attack operations—but the research makes a compelling case that defenders must shift from content-based assessment to behavior-based detection. Evaluating threats based on document quality indicators like unnatural translation, poor formatting, and spelling errors no longer works because AI excels at producing convincing decoys.
Beyond using indicators of compromise to spot attackers, organizations should watch for unusual behaviors following LNK file execution—including PowerShell activity, persistence mechanisms being established, and external communications—to hunt for threats. The report's core recommendation is that security teams can no longer rely on spotting amateurish language or formatting mistakes to flag phishing attempts, since AI-generated lures now mirror genuine business communications in structure and tone. The rise of state-sponsored hackers wielding advanced language models locally means defenders face adversaries who can iterate faster and produce more credible deception at scale. Enterprises relying on traditional email filtering may find their assumptions about what "looks suspicious" rapidly becoming obsolete, while security operations centers will need to prioritize behavioral anomaly detection over content inspection to keep pace.