An analysis found that Kimsuky, a North Korea–linked hacking group, used an "artificial intelligence (AI) coding agent" for hacking.
Genians Security Center (GSC), a domestic security firm, said on the 7th that it confirmed this after analyzing 13 malicious files collected on the previous month in a report.
Kimsuky distributed emails with malicious compressed files attached. The compressed files contained documents with names that looked legitimate, such as "insurance premium payment guide," "interest payment guide," and "policy fund guide." When a user runs these, the malware executes.
Genians said that, after analyzing the lure documents, it found traces indicating they were created with "opencode," an open-source AI coding agent.
The document contents also showed signs of AI use. In sections where specific figures such as payment cycles or interest rates should have been inserted, the phrase "(temporary value)" remained unedited. This is interpreted as the hacker distributing an AI-generated draft without proper review.
Earlier, Genians disclosed indications that Kimsuky used AI and a local large language model (LLM) for attack preparation and lure creation. However, this is the first time the use of an AI coding agent like opencode has been confirmed.
When the malicious file runs, it connects to a pre-created GitHub account to receive additional commands. Some variants also used Pastebin, a text-sharing site, as a channel to deliver commands. Even if access to GitHub is blocked, the communication path was dualized so that attack commands can continue to be received via Pastebin.
On infected computers, scheduled tasks disguised as BitLocker, a built-in Windows feature, or the software "MATLAB" are also covertly registered, and they run repeatedly at fixed intervals starting 5 minutes later to keep downloading additional malicious commands from GitHub.
The new variants have also strengthened features to evade detection. They first check whether tools used by security analysts are running and stop immediately if found. They also delete their own command execution records to erase analysis traces.