How are North Korean hackers using AI beyond drafting phishing emails?

They are building sophisticated local AI environments with tools like RAG for document search, AI agent frameworks, and AI-assisted coding to automate malware creation and analyze stolen intelligence.

What should organizations do to defend against AI-enabled cyber threats?

Researchers recommend moving from indicator-based detection to behavior-based detection, correlating sequences of anomalous activities like PowerShell execution and external communications.

North Korea's Cyber Arsenal Now Runs on Local AI

Q: What local AI tools did North Korean hackers use to evade monitoring?

The North Korean group Kimsuky used local AI tools Ollama, GPT4All, and Msty to process documents offline, preventing data from being flagged by commercial AI providers.

Q: Why is the use of local AI by hackers a strategic concern?

Using offline, open-source AI models removes the ability of companies like OpenAI to monitor and shut down malicious accounts, giving hackers permanent, untraceable AI access.

The Artillery Is Already Deployed

A key tactic discovered by Genians researchers is the use of local AI tools to evade detection. The North Korean group Kimsuky used Ollama, GPT4All, and Msty to process documents offline, preventing any sensitive data from being flagged by the monitoring systems of commercial AI providers.

This is not about simple automation. The hackers built a full-scale digital workshop, including:

Retrieval augmented generation (RAG) tools for searching stolen documents
AI agent development frameworks
Cursor, an AI-assisted coding tool

"What was observed in the threat actor's infrastructure was not merely evidence of several documents being created with AI, but a consistent process of capability development," Genians concluded.

The implication is clear. This goes far beyond a few polished phishing lures. It is operational infrastructure, enabling faster, more scalable, and more sophisticated attacks. This evolution mirrors a critical shift we've seen elsewhere, where attackers are now chasing machine speed on the show floor.

Mediocre Operators Now Have a Lethal Edge

The data shows this pivot is working. Another report by cybersecurity firm Expel details a separate North Korean operation, dubbed HexagonalRodent, that used ChatGPT and Cursor to "vibe code" almost an entire campaign. The result? An operation that installed malware on over 2,000 computers and stole an estimated $12 million in cryptocurrency in just three months.

Security researcher Marcus Hutchins, who discovered HexagonalRodent, told WIRED, "These operators don't have the skills to write code. They don't have the skills to set up infrastructure. AI is actually enabling them to do things that they otherwise just would not be able to do."

Evidence of AI authorship:

Code filled with English comments and emojis, unusual for North Korean programmers.
Fully AI-generated fake company websites used as phishing lures.
A large, exposed database tracking victim wallets, suggesting operational scale over sophistication.

This levels the playing field in a dangerous way. North Korea can now field larger teams of less-skilled operators, handing them AI models as a force multiplier. As Hutchins notes, North Korea has "hundreds of people being sent over the border to work in IT operations, and only a few of them really know what they're doing." AI provides the critical "leg up."

Why Local AI Is a Geopolitical Game-Changer

Using offline, open-source AI models is a calculated strategic decision with major implications.

It removes the primary choke point for defenders: the ability of companies like OpenAI to monitor and shut down malicious accounts. By operating locally, Kimsuky gains permanent, untraceable access to AI capabilities.

It embeds AI into the core attack lifecycle:

Previous Use (Limited)	New, AI-Integrated Capability (Kimsuky)
Drafting phishing email text	Automating full phishing campaign creation (sites, docs, lures)
Basic social engineering	Using RAG to intelligently query stolen data for intelligence
Manual vulnerability research	Using AI coding assistants to write and iterate malware

This creates a direct asymmetry. While democratic nations and their tech firms debate AI ethics and implement guardrails, state actors like North Korea operate with no such constraints. Their development cycle is faster, more secretive, and purely offensive. It’s a form of sanctions-proofing for cyber operations, building sovereign attack tools that cannot be easily taken away.

The Old Security Playbook Is Officially Obsolete

Genians' core recommendation is a strategic pivot: defenders must move from content-based assessment to behavior-based detection. Relying on known malware signatures (Indicators of Compromise, or IoCs) is no longer sufficient against AI-generated, constantly morphing code.

The new defensive posture must include:

Contextual correlation: Security systems need to stitch together sequences of anomalous activity, like a malicious LNK file execution, followed by unusual PowerShell commands, leading to persistent access, rather than just flagging a single bad file.
AI-powered defense: To counter AI-powered offense, defenders need their own machine-speed tools that can recognize novel attack patterns and adapt in real-time.
Assumption of automation: Security teams must now assume their adversaries can automate vast portions of the intrusion process, from initial access to data exfiltration.

This shift invalidates many traditional perimeter defenses. It demands a focus on identity, user behavior, and process integrity. The recent incident where hackers hijacked customer networks using 'God Mode' underscores how access, once gained, can be catastrophically exploited by automated tools.

What To Watch For Next

The trajectory points toward autonomous operations. The current use of AI agent development frameworks by Kimsuky is a clear stepping stone. The next phase will likely involve AI agents that can not only suggest attack methods but execute them, learn from the environment, and adapt tactics without human intervention.

The primary targets will evolve as well. While cryptocurrency theft remains a multibillion-dollar revenue stream, North Korean hackers stole over $2 billion in the first nine months of 2025 alone, the same AI capabilities are perfect for large-scale espionage. Theft of AI research and model poisoning could become strategic objectives.

The international response is lagging dangerously. Norms and treaties around cyber conflict were not built for an era of AI automation. The dilemma of regulating powerful open-source models, which fuel both innovation and weaponization, remains unsolved.

XOOMAR's analysis is that we are witnessing the professionalization of AI-enabled hacking by a state actor. This isn't a proof-of-concept. It is documented, operational, and financially successful. The barrier for entry has collapsed, turning middling IT workers into potent cyber operators. Defenders must now build systems that assume their adversaries have access to a limitless, automated, and learning toolkit. The alternative is to watch as attacks launched from the Hermit Kingdom achieve a scale and success rate once reserved for the world's most advanced cyber units.

Impact Analysis

States and individuals face a new class of automated, scalable cyberattacks that are harder to detect and counter.
The democratization of AI tools gives less sophisticated hacker groups a 'lethal edge,' raising the global threat level for all organizations.
The shift to local AI environments for operational tasks like malware creation makes traditional, signature-based defense systems increasingly obsolete.