"The more detailed the analysis, the better. Please do not do it haphazardly." Genians recovered that operator request on infrastructure it links to Kimsuky, underscoring a deliberate push to fold artificial intelligence into an espionage workflow rather than an accidental experiment.
Genians: offline AI tools on Kimsuky infrastructure
South Korean security firm Genians reports months of tracking and log analysis uncovered language-model tooling installed and run on servers it attributes to Kimsuky, a hacking unit under North Korea's Reconnaissance General Bureau. The firm found evidence of Ollama, GPT4All and Msty being used or configured, not merely downloaded: Ollama generated the keys created on first launch, and GPT4All carried a configured localdocs_v3.db — the database used by GPT4All's LocalDocs retrieval-augmented generation (RAG) feature.
Genians notes that it found no evidence the actor had trained its own AI model and characterizes the actor as in a "research and knowledge acquisition" stage. The presence of a LocalDocs database is specific evidence the operator tried to connect documents in its possession to an AI system, although Genians says the database alone does not establish those documents were stolen.
Developer libraries, transcription files and an AI coding editor
Beyond packaged apps, Genians recovered developer components that would let operators build AI functions into custom malware and tooling. The firm found libraries including LLaMaSharp, Microsoft's Semantic Kernel and Microsoft.Agents.AI for C# and .NET development, OpenAI's Whisper speech-to-text files together with a guide on extracting text from audio, and active traces of Cursor, an AI-powered coding editor. As Genians put it, "What is new is a nation-state espionage group assembling them on purpose to push AI deeper into its own attack workflow."
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleOperation GitPower: GitHub command channels, LNK-to-PowerShell and AsyncRAT
Genians ties the observed AI activity to an ongoing Kimsuky campaign it calls Operation GitPower. That operation abuses GitHub repositories as command channels within an LNK-to-PowerShell infection chain and has distributed encrypted AsyncRAT payloads disguised as image files. Separately, Fortinet documented a broader GitHub-C2 pattern in April that targeted South Korean users; Fortinet's work corroborates the surrounding technique family, while Reuters reported that the new local-AI artifacts could not be independently verified.
Genians' practical guidance for defenders shifts attention away from judging the quality of a lure toward detecting the intrusion: correlate LNK execution, PowerShell activity, hidden scheduled tasks, GitHub traffic and downstream payload behavior rather than relying on telltale spelling or formatting errors that AI-generated bait can eliminate.
How Genians ties this activity to Kimsuky and its prior operations
The firm links the activity to Kimsuky through overlaps with earlier campaigns, infrastructure clues and recovered operator logs containing North Korean vocabulary. The connection accords with the U.S. Treasury's 2023 designation that sanctioned Kimsuky and described the unit as subordinate to the Reconnaissance General Bureau and primarily focused on intelligence collection. Genians also notes a 2025 pattern it flagged when it linked Kimsuky to a spear-phishing attack that used ChatGPT-generated images of South Korean military employee ID cards.
What this means for technologists and security teams, policymakers and procurement leaders
- Technologists and security teams: Genians' findings suggest defenders must instrument and alert on execution and command-and-control signals — LNK launches, PowerShell chains, hidden scheduled tasks, unusual GitHub request patterns and later payload activity — because polished lures will no longer be a reliable indicator of legitimacy.
- Policymakers and regulators: The report documents a state actor assembling public and open-source AI tooling into an espionage workflow. That fact may influence assessments of how accessible components can be coordinated in nation-state operations and informs risk framing for regulatory discussion.
- Affected enterprises and procurement leaders: The evidence that an espionage unit is integrating RAG databases and offline language models highlights the need to inventory where local-model tooling, transcription tools and developer AI libraries are allowed on corporate infrastructure and to prioritize detection on behavior rather than message quality.
Genians' work paints a picture of groundwork being laid: packaged models, retrieval databases, transcription tools and development libraries assembled for integration into targeted operations. So far, the offline stack has not been shown running against a victim in reporting to date, and no GitPower victim count has been disclosed. The immediate operational implication is clear — with "nothing here to patch," defenders must shift to watching what attackers do after the bait is opened.
