North Korea's Hackers Are Running Language Models Offline

South Korean security firm Genians has published research showing that Kimsuky, a hacking group tied to North Korea's intelligence services, has used AI-generated documents in a pattern of spear-phishing attacks since 2026. Targets are in the military, diplomacy and academia.

One detail in the report deserves more attention than the headline got.

The models run locally

To avoid detection, Kimsuky has used open-source tooling — Ollama, GPT4All and Msty — to run large language models without an internet connection.

That single choice defeats most of what the industry currently calls AI safety.

Provider-side abuse detection. Usage monitoring. Account bans. Rate limits. Refusal training enforced at the API boundary. Every one of those controls assumes the model is reached over a network somebody operates and polices. A model running on a machine that never phones home answers to nobody.

The industry has put enormous effort into guardrails at the API layer. This is what it looks like when an adversary declines to use the API.

What the AI is doing

The attacks automate creation of malicious files disguised as legitimate documents — research reports, invitations, the ordinary paperwork of academic and diplomatic life.

Genians is precise about why that matters: "AI can generate highly polished documents on a wide range of topics within a short period of time, making it a highly efficient tool for threat actors." The firm adds that the significance runs past a change in how decoy documents get made, demonstrating that AI enables "the automation and large-scale production of social engineering attacks."

Craft has become manufacturing. A convincing spear-phishing lure aimed at a specialist used to need a human who understood the target's field well enough to write plausibly inside it. That was the bottleneck, and it kept targeted attacks expensive and therefore uncommon.

The bottleneck is gone. Plausible domain-specific documents can be produced at volume, across any field, by an operator who knows nothing about the subject matter.

Who this is

Kimsuky and other North Korean state-linked groups have been blamed for a long list of attacks, many aimed at financial gain. British blockchain analytics firm Elliptic reported North Korean hackers stole cryptocurrency worth more than $2 billion in the first nine months of 2025. In 2014, US authorities identified North Korea as responsible for the Sony Pictures hack.

Jenny Town, a senior fellow at the Stimson Center in Washington, DC, said the development was not surprising given the country's history, its hackers and programmers being more than capable of exploiting these tools.

Which is the right way to hold it. Not a new adversary — an existing, capable, well-resourced one whose cost per attack just fell.

What has to change on the defensive side

The uncomfortable implication: awareness training built around spotting bad grammar, odd formatting and generic phrasing now teaches people to look for a problem that has stopped existing in that form.

Detection has to move off linguistic quality and onto structural signals. Sender provenance. Authentication. Attachment behaviour. Whether the request makes sense given who is supposedly asking and why. Those hold regardless of how well the lure reads.

Start from the assumption that the message is flawless.

Source: North Korea's hackers using AI for attacks, cybersecurity firm says