North Korea's state-linked hacking group Kimsuky has reportedly found a way to harness artificial intelligence without relying on commercial services that could monitor or block their activities. According to South Korean cybersecurity firm Genians, the group is building its own AI tools to automate and sharpen its cyberattacks, using open-source models that run entirely on its own infrastructure.

The approach is notable for its deliberate evasion. Rather than using established chatbots such as ChatGPT or Claude, which typically log activity and enforce safety filters, Kimsuky is operating open models like Ollama, GPT4All, and Msty directly on its own machines. This keeps sensitive work away from any provider that might notice suspicious usage or report it to authorities.

A Deliberate Move to Stay Off the Grid

By keeping everything local, the group sidesteps the monitoring and safety mechanisms that companies like OpenAI have built specifically to catch malicious misuse. The strategy reflects a broader understanding of the AI landscape: publicly available software can be transformed into private weaponry, invisible to the outside world.

This is not merely a technical preference. It is an operational choice that reduces exposure to intelligence-gathering and legal consequences. Commercial AI services are rarely designed with anonymity in mind. They require accounts, track usage patterns, and often have terms of service that prohibit cyberattacks. For a state-sponsored espionage group, using those services would be reckless. Running local models removes the middleman and places all control in the hands of the operator.

A Toolkit Beyond Chatbots

The AI toolkit attributed to Kimsuky goes far beyond a simple chatbot interface. Genians reports that the group has integrated retrieval-augmented generation, commonly known as RAG, to sift through stolen documents and extract useful intelligence. This technique allows the AI to draw on specific databases of stolen information and generate responses grounded in that data, making analysis faster and more focused.

In addition, the group has adopted AI agent frameworks that can string together multiple tasks automatically. These frameworks enable a sequence of actions, such as searching for a file, summarizing its contents, and preparing a phishing email, all without manual intervention. Speech-to-text software has also been incorporated, likely to process voice recordings or audio intercepts. Perhaps most concerning is the reported use of Cursor, an AI-assisted coding tool, to write and refine malware. That would allow developers to iterate more quickly on malicious code and adapt it to evade security defenses.

More Convincing Deception

The integration of these tools has translated into more polished and credible social engineering campaigns. The researchers describe a notable improvement in the quality of phishing lures. Among the group's outputs are finance- and cryptocurrency-themed documents designed to mimic legitimate workplace reports. These are far harder for victims to dismiss than the poorly worded and obviously fraudulent phishing messages of past years.

By using AI to automate and refine each stage of an attack, Kimsuky can operate at a scale and speed that would be difficult to achieve with manual labor. Phishing emails can be tailored to specific targets based on stolen data, malware can be updated more rapidly, and stolen information can be analyzed much more quickly. The result is a more effective cyber-espionage operation that can compromise more victims in less time.

A State-Sanctioned Cyber Enterprise

Kimsuky is not an unknown entity to the international community. The United States Treasury Department blacklisted the group in 2023, describing it as a North Korean government-controlled cyber-espionage group that gathers intelligence to serve Pyongyang's strategic aims. The designation freeze any US assets tied to the group and prohibits American companies from doing business with it, but it does little to stop its underlying activities.

The group's behavior fits into a well-documented pattern of North Korean cyber operations targeting both secrets and money. Pyongyang's hacking crews have repeatedly gone after developers and cryptocurrency users. One notable method has involved distributing poisoned code packages on open-source repositories, designed to steal developer credentials and gain access to larger systems. The use of AI simply makes these campaigns cheaper to run at scale.

North Korean hacking groups are effectively state enterprises. Their operations are not opportunistic. Washington has sanctioned multiple North Korean groups responsible for major attacks, including the notorious WannaCry ransomware outbreak of 2017. That attack affected hundreds of thousands of computers across 150 countries and caused billions of dollars in damages. For North Korea, a heavily isolated economy in need of hard currency, cybercrime has become a central source of revenue.

The Broader Implications

What makes the AI angle especially worrying is how it lowers the cost of scaling operations. The same automation is also fueling a wider surge in online crime. According to recent estimates, the global scam economy has passed $442 billion. A well-resourced state actor is far better positioned to exploit these tools than a lone fraudster, meaning the impact could be disproportionately large.

There is also a thorny question of accountability. As autonomous software takes on more of the work in cyberattacks, it remains unclear who should be held responsible when an AI agent causes harm. Traditional legal frameworks assume human actors are making deliberate choices. With AI, the decision-making process is distributed across algorithms and training data. A hostile government running its own models sits well outside any provider's reach, making oversight even more difficult.

An Unverified but Plausible Report

One caveat is worth keeping in view. The findings from Genians have not been independently verified. The precise scope of the operation rests on a single firm's research, and without broader confirmation, it is still possible that some details are inaccurate or exaggerated. However, the report aligns with everything known about Kimsuky's capabilities and tactics. The group has long been recognized as one of North Korea's most active and sophisticated hacking units, with a track record of adapting new technologies to its operations.

The report also complicates the argument that open-source AI models are harmless by default. The very openness that allows researchers, startups, and independent developers to build freely is the same openness that allows a sanctioned state to run identical software beyond anyone's reach. This is a trade-off that the AI industry has been reluctant to confront head-on.

The broader lesson is uncomfortable all the same. The safety controls that leading AI labs tout publicly work only when the misuse runs through their own systems. Choosing open models on private hardware allows the world's most determined attackers to find an obvious way around those controls. As long as powerful AI tools remain freely available, it will be impossible to fully prevent malicious actors from weaponizing them. That is not an argument against open source, but it is a reminder that the benefits of open development come with real and persistent risks.

Source: TNW | Security News