North Korea's Kimsuky Group Develops Offline AI Stack for Advanced Cyber Operations
North Korean state-sponsored hackers, Kimsuky, are building an offline artificial intelligence stack to enhance phishing campaigns and automate malware development, moving away from public AI platforms.
North Korea's state-sponsored hacking collective, Kimsuky, has reportedly begun developing and utilizing an offline artificial intelligence (AI) stack, signaling a significant advancement in their cyber espionage capabilities. This move away from public AI platforms like ChatGPT indicates a strategic shift towards more sophisticated, self-sufficient, and potentially harder-to-detect cyber operations.
Security researchers at Genians uncovered evidence of this development while tracking infrastructure associated with Kimsuky, a group known to be subordinate to North Korea's Reconnaissance General Bureau. The analysis revealed that the group is assembling and testing existing AI tools rather than training novel models. The apparent goal is to integrate AI capabilities across their entire attack workflow, from crafting more convincing phishing lures to analyzing intelligence data and automating malware development.
Key components of this offline stack include tools for running large language models locally, such as Ollama, GPT4All, and Msty. Genians found evidence that these tools were not merely downloaded but actively configured and used. For instance, GPT4All was found with a configured LocalDocs database, indicating the group's intent to use retrieval-augmented generation (RAG) to enable AI models to answer queries based on their private document collections. While there's no confirmation that these documents were stolen, the setup suggests an effort to leverage internal data with AI.
Beyond language models, Kimsuky appears to be gathering developer libraries and tools necessary for building custom AI functions into their malware. Found on the same infrastructure were components like LLaMaSharp and Microsoft's Semantic Kernel, which are used for creating AI-powered C# and .NET applications. Additionally, traces of OpenAI's Whisper speech-to-text technology and guides for audio transcription were discovered, alongside the AI-powered coding editor Cursor.
This development is particularly concerning for defenders. AI-generated phishing lures can bypass traditional detection methods that rely on linguistic errors or formatting inconsistencies. Genians advises security teams to focus on correlating low-level execution artifacts, such as LNK file execution, PowerShell activity, hidden scheduled tasks, and GitHub traffic, rather than solely relying on the perceived quality of a phishing email.
The newly observed offline AI stack is reportedly being integrated into Kimsuky's ongoing Operation GitPower campaign. This campaign has previously been observed abusing GitHub repositories as command-and-control channels and distributing encrypted AsyncRAT payloads disguised as image files. While the specific use of the local AI stack against victims has not yet been demonstrated, the groundwork for automating various stages of their operations is clearly being laid.
Kimsuky's previous activities, including a 2025 spear-phishing attack that used ChatGPT-generated images of South Korean military ID cards, demonstrate a history of adopting new technologies for espionage. The U.S. Treasury sanctioned Kimsuky in 2023, identifying it as a key intelligence-gathering unit for the North Korean regime. This latest move towards an offline AI infrastructure represents a significant escalation in their operational sophistication and a growing challenge for global cybersecurity defenses.
The implications of state-sponsored actors developing self-contained AI capabilities are far-reaching. It suggests a future where cyberattacks could be faster, more personalized, and more difficult to attribute, requiring a fundamental shift in defensive strategies and threat intelligence gathering.