North Korean Hackers Build Offline AI Stack to Automate Cyberattacks and Evade Detection

North Korean Hackers Deploy Local AI Infrastructure

North Korean hackers are no longer relying on public chatbots for their operations. Kimsuky, a cyber-espionage unit operating under North Korea's Reconnaissance General Bureau, has built its own offline AI infrastructure to automate and enhance cyberattacks, according to South Korean security firm Genians1

1

2

2

. The state-sponsored hackers are running local large language models using Ollama, GPT4All, and Msty on their own servers, keeping all activity away from commercial AI providers that monitor for abuse3

3

. This strategic choice allows the group to process sensitive information without external exposure, making it particularly attractive for a state-sponsored threat actor conducting intelligence operations.

AI Tools for Cyberattacks Assembled Over Months

Genians conducted months of tracking and log analysis on infrastructure tied to Kimsuky, uncovering an extensive collection of AI tools for cyberattacks4

4

. The researchers found evidence that the group configured Ollama to generate keys on first launch and set up GPT4All's LocalDocs retrieval-augmented generation feature with a configured database2

2

. Beyond ready-made applications, Kimsuky collected developer libraries including LLaMaSharp, Microsoft's Semantic Kernel, and Microsoft.Agents.AI—components for building AI functions into custom software1

1

. The group also gathered OpenAI's Whisper speech-to-text models and actively used Cursor, an AI-assisted coding tool, to edit code1

1

4

4

. The development components spanning local AI execution, document retrieval through retrieval-augmented generation, automated agents, and external AI integration strongly suggest these tools were collected for direct development of AI-based attack capabilities.

Phishing and Malware Development Enhanced by Open-Source AI Models

Kimsuky is integrating AI throughout its attack chain, from phishing and malware development to data analysis. The group's recent phishing campaigns use ZIP archives containing malicious LNK files disguised as materials related to international events, research reports, or meeting requests1

1

. When victims execute these files, an embedded PowerShell loader collects extensive system information to assess the infected environment. Genians found that AI-enabled threats now include finance and cryptocurrency-themed decoy documents generated with AI that use natural language, polished structure, and formats similar to actual business materials to increase user trust1

1

5

5

. The attackers employ various obfuscation techniques, including Base64 encoding, string splitting, and custom decoding routines to hide malicious behavior1

1

.

Command-and-Control Infrastructure Reveals AI Research

The investigation uncovered that Kimsuky continues using Git repositories for command-and-control infrastructure, with multiple public GitHub repositories containing configuration files, PowerShell scripts, and various payloads1

1

. These repositories also serve as testing grounds for malware development, stolen data management, and AI technology research. Researchers recovered operator logs showing requests to analyze data sets for wallet details, Gmail credentials, and site-registration history, with instructions stating "The more detailed the analysis, the better. Please do not do it haphazardly"2

2

. The use of retrieval-augmented generation on stolen files can help attackers more quickly and automatically identify valuable information within large volumes of data1

1

.

Offline AI Infrastructure Bypasses Commercial Safeguards

By running open-source AI models on private hardware, North Korean hackers have found a way around the safety controls that leading AI labs promote3

3

. Commercial AI providers like OpenAI have built monitoring systems and safety filters to catch misuse, recently terminating multiple ChatGPT accounts used in phishing and human trafficking4

4

. However, these safeguards only work when misuse runs through them. The offline AI infrastructure allows Kimsuky to harness AI capabilities without ever handing their secrets to a Silicon Valley company, turning freely available AI software into an in-house weapon nobody else can see running3

3

. The US Treasury sanctioned Kimsuky in 2023 as a North Korean government-controlled cyber-espionage group that gathers intelligence to serve Pyongyang's strategic objectives2

2

5

5

.

Behavior-Based Threat Detection Becomes Critical

Genians emphasized that defenders must shift from content-based assessment to behavior-based threat detection as AI makes traditional indicators unreliable1

1

4

4

. Assessing threats based on the quality of fake documents—such as unnatural translated language, poor formatting, and spelling errors—is no longer effective because AI excels at producing convincing decoy documents. Organizations should look for anomalous behaviors following LNK execution, such as PowerShell execution, persistence establishment, and external communications, to assess overall threat levels1

1

4

4

. While Genians found no evidence that North Korean hackers have begun training their own models, the findings demonstrate they are in a research and knowledge acquisition stage, assembling and testing existing tools with the apparent aim of folding AI throughout their operations2

2

. Watch for increased sophistication in state-sponsored attacks as AI tools for cyberattacks become more integrated into hostile operations, and expect defenders to rely more heavily on behavioral analysis rather than content quality to identify threats.