North Korea’s state-sponsored hackers have found a way to harness artificial intelligence without ever handing their secrets to a Silicon Valley company. According to South Korean cybersecurity firm Genians, the state-linked group known as Kimsuky has been building its own AI tools to automate and sharpen its cyberattacks. The operation is a stark demonstration of how open-source AI can be weaponized by determined adversaries.

The most deliberate part of the strategy is how carefully the group stays off the grid. Rather than rely on commercial chatbots, which typically log activity and screen for abuse, Kimsuky is running open models such as Ollama, GPT4All and Msty directly on its own machines. This local approach keeps sensitive work away from any provider that might notice or report it, turning freely available software into a hidden in-house arsenal.

That choice is the whole strategy in miniature. By keeping everything local, Kimsuky sidesteps the monitoring and safety filters that firms like OpenAI have built precisely to catch this kind of misuse. Commercial AI platforms are designed to refuse harmful requests and flag suspicious activity, but self-hosted open models can be modified or used with no such restrictions. The group therefore gains the full power of modern AI without exposing its operations to outside oversight.

A covert AI pipeline

According to Genians, the toolkit goes well beyond a single chatbot. Kimsuky has reportedly pulled in retrieval-augmented generation, commonly known as RAG, to sift through stolen documents. RAG allows an AI system to query a database of documents and generate answers based on that material, which is ideal for an espionage group that collects large volumes of confidential files. The same technique is used by legitimate enterprises to build internal knowledge assistants, but in the hands of hackers it becomes a way to rapidly extract intelligence from loot.

The group is also said to be using AI agent frameworks, which chain multiple tasks together into a single automated workflow. Instead of manually moving from one step to the next, an attacker can define a goal and let the agent plan and execute the required actions. For a cyber operation, that might mean scanning targets, harvesting credentials, and sending spear-phishing emails with minimal human intervention. This is a significant leap forward from the largely manual tradecraft that characterized state-sponsored hacking for years.

Genians researchers further identified speech-to-text software as part of the arsenal. This could be used to process audio intercepted from phones, meetings, or other surveillance sources, converting it into searchable text for analysis. It may also help in building more convincing social engineering attacks, such as voice phishing or voicemail impersonation, by giving attackers a better understanding of language patterns used by their targets.

Perhaps most worryingly, the group has adopted Cursor, an AI-assisted coding tool, to help write and refine malware. Cursor is a popular development environment that uses large language models to autocomplete code, explain snippets, and suggest fixes. While it is designed to accelerate legitimate software development, it works just as well for writing malicious code. With Cursor, even moderately skilled developers can produce polished and effective malware far faster than they could by writing every line from scratch.

More believable phishing, faster malware

With that stack in hand, Kimsuky can work faster and more convincingly across the board. The researchers describe a group that is automating attacks, crafting more believable phishing lures, weaving AI into malware development, and analyzing whatever it steals. The days of clumsily worded scam emails appear to be giving way to a more dangerous, AI-polished brand of social engineering.

The deception has indeed become notably slicker as a result. Among the group’s outputs are finance- and cryptocurrency-themed documents designed to mimic legitimate workplace reports. These are the sort of professional-looking baits that are far harder to dismiss than the phishing attempts of years past. A target who receives a well-formatted document that matches internal templates is far more likely to open it, and with a single click the attack chain begins.

AI also enables scale. A single attacker can generate hundreds or thousands of highly personalized messages in minutes, a task that once required a large human team. Those messages can be automatically translated into multiple languages, customized with personal details gleaned from social media or stolen correspondence, and even checked for grammar and tone consistency. For a group responsible for conducting espionage on behalf of the North Korean regime, this is a powerful force multiplier.

A familiar group, evolving tradecraft

None of this comes out of nowhere. Kimsuky is a known and sanctioned operator in the cyber threat landscape. The US Treasury blacklisted the group back in 2023, describing it as a North Korean government-controlled cyber-espionage group that gathers intelligence to serve Pyongyang’s strategic aims. Kimsuky is also tracked under names such as APT43, Emerald Sleet, and Nickel Kimball by various cybersecurity firms.

The group has been active for more than a decade, focusing largely on government agencies, academic institutions, and think tanks that deal with foreign policy and security issues. Its long-running campaigns have often involved spear phishing against South Korean targets, as well as against experts and policymakers in the United States. The group typically tries to harvest credentials, gain access to email accounts, and exfiltrate sensitive research and analysis.

The new AI capabilities do not suggest a complete break from the past. Rather, they are an evolution of a proven playbook. Kimsuky’s operators are simply adding modern tools to a well-refined set of techniques, many of which have been publicly documented by researchers over the years. The United Nations and national cybersecurity agencies have repeatedly called out North Korean hacking groups for using cybercrime to raise money and blunt sanctions.

Pyongyang’s criminal digital operation is vast. North Korean crews have repeatedly targeted developers and cryptocurrency users, using techniques such as poisoned code packages designed to steal developer credentials. There have been numerous reported incidents of fake job interviews, malicious NFT links, and bogus cryptocurrency trading apps, all aimed at eliminating the financial isolation imposed by the international community.

Washington has sanctioned several North Korean groups behind attacks such as WannaCry, the global ransomware worm that shut down hospitals and businesses in 2017. Those sanctions reflect how central cybercrime has become to a heavily isolated economy in need of hard currency. For Pyongyang, hacking is not simply a side project; it is a state-backed enterprise with clear strategic and financial objectives.

A lower cost of attack at scale

What makes the AI angle particularly worrying is how sharply it lowers the cost of scale. In the past, running a large espionage campaign required considerable manpower, careful coordination, and time. AI automation collapses many of those costs. A single operator can now manage multiple campaigns, adapt quickly to new targets, and continuously refine malware, all while spending fewer resources than before.

The same automation is helping fuel a wider surge in online crime. The global scam economy is estimated to have passed $442 billion, according to recent analyses of online fraud and cybercrime. While the bulk of that activity is carried out by ordinary criminals, a well-resourced state actor is far better placed to exploit these tools than a lone fraudster. The combination of state funding, technical expertise, and advanced AI models creates a threat that is qualitatively different from what came before.

There is also a broader accountability question that the technology industry has yet to resolve. As autonomous software takes on more of the work, it remains unsettled who is responsible when an AI agent causes harm. Legal systems are still built around human actors making deliberate choices. When a malware campaign is planned and executed with the help of an autonomous agent, the chain of responsibility becomes murkier. And a hostile government running its own models sits well outside any provider’s reach in the first place.

The open-source dilemma

One caveat is worth keeping in view. Genians’ findings could not be independently verified, so the precise scope of the operation rests on a single firm’s research. Cyber threat reports often contain uncertainties, and it is possible that some details are inaccurate or inferred. Nevertheless, the report aligns with everything known about Kimsuky’s historical behavior and broader trends in the use of AI by malicious actors.

The findings also complicate the comfortable assumption that open models are harmless by default. Open-source AI has long been celebrated as a democratic force, enabling researchers, startups, and individuals to build applications without waiting for a few large companies to grant access. That openness is a genuine virtue, but it is a double-edged sword. The very same openness that lets a small team create a useful medical assistant is what lets a sanctioned state run the same software beyond anyone’s reach.

Model developers have spent considerable effort trying to bake safety filters into their systems, but those filters are only as strong as the deployment environment. Once an open-weight model is downloaded and running on private hardware, the operator can strip out safety prompts, fine-tune the model on malicious data, or use uncensored variants that refuse nothing. No amount of red-team testing or responsible disclosure can fully prevent this, because the code is simply out there.

The broader lesson is uncomfortable all the same. The safety controls that leading labs tout work only when the misuse runs through them. By choosing open models on private hardware, the world’s most determined attackers have found the obvious way around them. The same story has played out again and again across the history of technology: from encryption to dual-use software, tools that empower everyday people also empower those with the most harmful intentions. The AI industry is only beginning to grapple with that reality.

For defenders, the implication is clear. They cannot assume that the largest AI providers will catch every malicious request, because the attackers are no longer using those services. Instead, cybersecurity teams must treat AI as a baseline assumption in their threat models, preparing for adversaries who have the same level of automation and intelligence that they do. The challenge is no longer theoretical. It is already being deployed in phishing kits, malware development pipelines, and intelligence operations around the world.

As the boundaries between human and machine decision-making continue to blur, the intelligence community will have to adapt. The story of North Korea’s hackers building their own AI tools is a reminder that every beneficial technology eventually finds its way into hostile hands. How well the world manages that reality will shape the security landscape for years to come.

Source: TNW | Security News