Kimsuky, the North Korean state-sponsored group, has moved beyond testing public AI APIs and is now engineering AI into its attack infrastructure. The group runs offline language models (Ollama, GPT4All, Msty) on command-and-control servers to process stolen documents without exposing operations to cloud providers. Genians researchers also found development frameworks including LLaMaSharp and Microsoft.SemanticKernel, suggesting this is no longer experimentation but systematic deployment across malware development, phishing document generation, and stolen-data analysis.
The infrastructure choice is revealing. Offline inference eliminates API logs that cloud providers could detect and that defenders could alert on. For a group running hundreds of attacks monthly, external dependencies become unacceptable risk. They’re also routing command-and-control through GitHub repositories with AsyncRAT payloads disguised as images.
This is operational hardening.
Sources: Genians Security Center threat intelligence report
Coverage: The Hacker News • The Block
Related on this blog: The Open Weights Letter Grew Ten Signatures Overnight, and One of Them Was OpenAI • Google Signed the Open Weights Letter, and Anthropic Is Now Alone • JADEPUFFER’s 19-Day Upgrade: Ransomware Built to Destroy AI Models