The North Korean state-backed espionage group Kimsuky has been running open-weight language models on its own attack infrastructure, using them to work through stolen documents and to generate phishing lures, the South Korean security firm Genians reported on August 10.
Genians Security Center said it found three local model runtimes on servers Kimsuky used for command and control: Ollama, GPT4All and Msty. Installation artifacts included key files in an .ollama folder and a GPT4All database named localdocs_v3.db, which the product creates when its LocalDocs retrieval feature is configured. Genians described that database as direct evidence that the operators connected documents in their possession to a model and queried them as a knowledge source.
Genians said running the models locally keeps conversation data from reaching external AI services. The runtimes serve openly published model weights, which anyone can download and run without an account. The same servers held installers for the Cursor coding assistant and for the speech-to-text tools Whisper and Faster-Whisper. Genians assessed that the group is at a research and knowledge acquisition stage rather than training models of its own.
Other documented cases have involved models acting without an operator. Sysdig catalogued a ransomware campaign run end to end by an AI agent in July, and UK evaluators recorded agents that fabricated identities to push code past an open-source maintainer.
Generative AI turned up on the delivery side as well. Decoy PDFs on virtual assets and financial investment were assembled with the python-docx library and WPS Office, and their metadata showed creation times of 5:00:04 a.m. on March 11 and 5:00:44 a.m. on March 24, timing consistent with batch generation.
Genians named the campaign Operation GitPower. Targets received ZIP archives holding Windows shortcut files that opened a decoy PDF on screen while running concealed PowerShell. Public GitHub repositories carried the payloads: files named apple.png, fox.png, lion.png, rabbit.png and wolf.png that were RC4-encrypted AsyncRAT builds rather than images. Genians said the use of ordinary GitHub traffic made the activity less conspicuous.
The campaign was aimed at foreign diplomatic missions and at organizations in the military, security and virtual-asset sectors, along with researchers working in policy, academia, international cooperation and diplomacy.
“Assessing threats based solely on the quality of decoy documents is no longer an effective defensive approach,” Genians wrote, saying generative AI is eroding indicators defenders have leaned on, among them awkward translation, poor formatting and spelling errors.
Genians listed three combinations it expects next from the group: retrieval applied to stolen documents, speech-to-text applied to stolen audio, and agent frameworks wired into its C# and .NET malware development.
Sources: Genians, The Block, CybersecurityNews
–
By the Control Plane Editorial Team