North Korea's Kimsuky hackers are running a full local AI stack, Genians reports

North Korean hackers are running AI locally now, and the location is the whole point.

What Genians found

On August 10, 2026, South Korean security firm Genians reported that infrastructure tied to the North Korean group Kimsuky was carrying a full local AI stack. The inventory reads like a hobbyist’s setup guide: Ollama, GPT4All and Msty for running and managing models on their own machines, retrieval augmented generation for searching documents, AI agent development frameworks, speech to text software, and the AI coding tool Cursor.

Kimsuky is not a new name. The US Treasury sanctioned the group in 2023 as a North Korean government-controlled cyber-espionage operation.

Why local matters

Running models on your own hardware means stolen documents can be processed without ever being sent to an outside AI service, where the activity might be logged, refused or flagged. Every safety mechanism the commercial AI providers have built, usage monitoring, account bans, abuse classifiers, sits on the server side. A local model has no server side.

Genians’ conclusion, in its own framing: the findings suggest Kimsuky is moving beyond using generative AI to create phishing lures and is building capacity to integrate existing AI models into malware development, data analysis and attack automation. The RAG tooling in particular fits an espionage workflow, since it is exactly what you would use to search and summarize a large pile of exfiltrated documents.

The open weights problem, made concrete

The detail worth sitting with is how ordinary the tool list is. Ollama, GPT4All, Cursor: this is the same stack installed on a large share of developer laptops. There is no exotic capability here, no bespoke model, just publicly available software assembled the way any competent engineer would assemble it.

That is the open weights debate in miniature. Open models do not check who is running them, and the guardrails that exist for hosted AI simply do not apply once the weights are on a sanctioned group’s own machines. The Genians report does not resolve where the line on open models should sit, but it removes the hypothetical: state-linked operators are already on the other side of it.