Microsoft: North Korean Hackers Using AI to Mass-Distribute 100

Microsoft's investigation has confirmed that North Korean hacking groups used generative artificial intelligence (AI) to churn out dozens of customized phishing lure documents targeting South Korean users in a matter of minutes. Warnings are mounting that as precisely tailored bait documents are mass-produced, detection and blocking will become increasingly difficult.

According to a report on North Korean hacking groups' AI usage published by Microsoft Threat Intelligence (MSTIC) on the 29th (local time), the hacking group "Onyx Sleet" (known in South Korea as Andariel), believed to be affiliated with North Korea's Reconnaissance General Bureau, produced more than 100 PDF lure documents targeting South Korean users in August and September of this year. By analyzing metadata such as file creation timestamps, Microsoft detected traces of dozens of documents being created in succession within minutes, and concluded on that basis that it is highly likely an AI platform was used. It did not disclose how it obtained the files.

The concern is that these phishing documents were produced at a speed unlike anything seen before. In the past, creating a single lure document tailored to a target's job title and interests took weeks; now AI generates dozens in a row in just minutes. With documents able to be quickly regenerated and revised, critics say security firms' detection and blocking systems are effectively left to respond only after the fact. However, it has not been confirmed whether the South Korea-targeted documents were actually sent or how much damage they caused.

AI use is spreading beyond phishing document production to actual intrusion operations and malware development. Microsoft confirmed that since late October, Onyx Sleet and "Pearl Sleet" have used ARTEX, an open-source agentic AI penetration-testing framework, in cyber operations. The approach has shifted from hackers manually hunting for server vulnerabilities one by one to a structure in which, once a target address is entered, AI analyzes security flaws, designs an intrusion path, and even generates executable code.

AI has also been deployed in malware development. Since last April, "Coral Sleet" has introduced into actual operations an attack procedure that inserts malicious code into source code in private GitHub repositories belonging to companies and others. Microsoft assessed that the feature's rapid development and widespread use make it highly likely that AI-based development tools were used. Another group, "Sapphire Sleet," is also suspected of using AI to improve malware capabilities in an attack targeting software developers last August. However, the targeted companies and the scale of damage were not disclosed.

Microsoft classified the AI adoption levels of eight major North Korean hacking groups it tracks into four tiers. Three—"Jade Sleet," Coral Sleet, and "Jasper Sleet"—were rated as having reached the highest level, "full integration." Four—"Diamond Sleet," Onyx Sleet, Pearl Sleet, and Sapphire Sleet—were classified at the "accelerating" stage, where use is spreading rapidly, while "Emerald Sleet" was placed at the "expanding" stage, using AI for some tasks. No group remained at the initial experimental level.

Note: Results of Microsoft's four-tier classification of AI usage levels across eight North Korean hacking groups

Still, the attacks are not fully autonomous. Microsoft analyzed that while AI supports reconnaissance, malware development, and automation of repetitive tasks, target selection and key decision-making remain human-led. North Korean groups' AI adoption is likewise described as focused on raising the speed and efficiency of operations while maintaining their existing goals of generating revenue and collecting strategic intelligence.

The implication of the report is that the "productivity" of cyber threats itself has changed. With the ability to prepare large volumes of tailored attacks, corporate security teams can no longer defend themselves by filtering the signatures of individual malicious files alone. In particular, techniques that directly target source code repositories can contaminate an entire supply chain, prompting calls for heightened vigilance among South Korea's IT, gaming, and financial sectors, where development organizations are concentrated. Within the security industry, voices are growing that offsetting the speed of AI-driven attacks will require raising both detection automation and the embedding of security into the development process.