North Korean hackers stole $12 million in 90 days using AI—and security teams barely noticed

A group of North Korean hackers stole as much as $12 million in just three months—using artificial intelligence tools to do work that would have taken a skilled cybercriminal team months to execute manually. The speed and scale of the theft exposed a critical vulnerability in corporate security: generative AI is lowering the barrier to entry for financially motivated attackers with minimal technical expertise.

The theft matters because it signals a fundamental shift in the economics of cybercrime. When AI can automate the grunt work of hacking—writing malware, crafting social engineering attacks, building convincing fake infrastructure—even mediocre attackers become dangerous. Security teams designed to catch sophisticated adversaries may miss the noise generated by AI-assisted amateurs operating at scale.

According to reporting from Wired, the North Korean group weaponized AI across the entire attack chain. They used generative AI to write malware code, a technique sometimes called “vibe coding” where attackers feed AI systems prompts to generate functional exploit code without hand-crafting each line. They also deployed AI to create fake company websites designed to trick targets into revealing credentials or downloading malicious files. The combination allowed them to move from initial reconnaissance to theft in a compressed timeline that would have been impossible using traditional manual methods.

The $12 million figure represents confirmed or highly probable theft over a 90-day window. The actual scope may be larger, but attribution and damage assessment in cross-border cybercrime cases often lags months or years behind the actual incident. What made this case notable was not the final dollar amount—cryptocurrency theft and business email compromise schemes regularly exceed this—but the speed and the minimal human labor required to execute it.

How Did AI Transform Amateur Hackers Into Million-Dollar Thieves?

The hackers’ relative lack of sophistication is the critical detail. These were not elite nation-state operators with decades of institutional knowledge. They were attackers who, by traditional standards, would lack the skills to mount a campaign of this scale and speed. AI closed that gap. A generative AI model trained on public code repositories and security research can produce working malware variants. The same tools can generate convincing phishing emails and fake login pages that pass initial visual inspection. What once required specialized knowledge became a prompt-engineering problem.

CISA’s threat intelligence documents how North Korean APT groups have evolved their tactics, but the integration of AI tools represents a significant acceleration in their capabilities. This shift toward AI-powered attack methods mirrors broader trends in how artificial intelligence is being weaponized across multiple threat vectors.

Why Are Traditional Security Systems Missing AI-Generated Attacks?

Security teams at targeted organizations reportedly failed to detect the breach during the 90-day window, which underscores a second risk: defensive tools and processes were not calibrated to flag AI-generated attack patterns. Traditional intrusion detection systems look for signatures—known malware hashes, familiar command sequences, recognized attacker infrastructure. AI-generated malware produces novel signatures on each iteration. Fake websites built by generative design tools may not trigger the same red flags as manually crafted phishing pages. The defenders were looking for patterns that no longer applied.

No major company has publicly acknowledged being the victim in this specific case, and Wired’s reporting does not name a target. That opacity is itself a problem: victims of ransomware and theft often delay disclosure for weeks or months while assessing damage and negotiating with insurers. By the time public disclosure occurs, the attacker has already moved funds through cryptocurrency mixers and established new infrastructure. The speed advantage remains with the attacker.

Are AI Safety Guardrails Actually Working?

The incident also raises questions about the security practices of AI tool providers. Generative AI platforms like ChatGPT, Claude, and open-source models like Llama have implemented safeguards to refuse requests for malware code or social engineering templates. But determined users can work around these guardrails through prompt injection, jailbreaking, or by using less-restricted open-source models. The North Korean group’s success suggests that either these safeguards are insufficient or that attackers have found reliable methods to bypass them.

The implications extend beyond traditional cybersecurity into the realm of synthetic media creation, where similar AI tools can generate convincing fake content for social engineering attacks. This convergence of capabilities means that attackers can now create both the technical infrastructure and the persuasive content needed for comprehensive campaigns.

What Should Organizations Do Right Now?

For organizations, the immediate implication is clear: security teams need to assume that attackers now have access to AI-assisted capabilities and that traditional detection methods will miss some attacks. That means shifting focus toward behavioral analysis, network segmentation, and credential security—defenses that work regardless of whether the malware was hand-coded or AI-generated. It also means treating AI-generated phishing and social engineering as a higher-confidence threat, since the volume and quality of such attacks will only increase.

The defensive strategy must account for the reality that AI manipulation techniques are becoming more sophisticated and accessible. Organizations can no longer rely solely on employee training to spot obvious phishing attempts when AI can generate highly personalized and contextually appropriate attack content.

The broader question is whether cybersecurity can scale fast enough to match the pace that AI has introduced to offensive operations. If a small group of mediocre hackers can steal $12 million in 90 days using publicly available AI tools, what becomes possible when better-resourced criminal organizations or nation-states apply the same techniques at full capacity? The answer will likely shape corporate security budgets and regulatory pressure on both AI companies and enterprise security vendors for years to come.