Kimsuky APT Uses ChatGPT to Generate Fake South Korean Military IDs for Espionage Campaign (2025)

Incident Summary

North Korean advanced persistent threat (APT) group Kimsuky successfully used ChatGPT to generate fake South Korean military identification documents by framing requests as “sample designs,” bypassing OpenAI’s content safeguards through social engineering of the AI system itself.^[1] The fabricated military IDs were deployed as part of an espionage campaign targeting researchers specializing in North Korea studies, using the convincing documents to establish false credibility in phishing and social engineering operations.^[2][3] The incident demonstrates that nation-state threat actors have developed techniques for bypassing AI safety guardrails not through technical jailbreaks but through the simpler approach of reframing malicious requests as innocuous design tasks. The “sample designs” framing exploited the ambiguity between legitimate design work and document forgery, a boundary that AI content policies struggle to enforce because the same output can serve either purpose depending on the user’s intent.

Key Facts

• Threat actor: Kimsuky APT (North Korean state-sponsored)^[1]
• Method: Framed requests as “sample designs” to bypass safeguards^[1]
• Output: Fake South Korean military identification documents^[1]
• Use: Espionage campaign targeting NK studies researchers^[3]
• AI system: ChatGPT^[1]

Threat Patterns Involved

Primary: Social Engineering via AI — Kimsuky’s use of ChatGPT to generate fake military IDs represents AI-assisted social engineering at two levels: the APT socially engineered the AI system itself (through framing), and then used the AI-generated documents to socially engineer human targets (researchers).

Significance

1. Nation-state AI exploitation — Kimsuky’s use of ChatGPT for document forgery confirms that nation-state APTs are integrating commercially available AI tools into their operational tradecraft
2. Framing bypass technique — The “sample designs” framing demonstrates a category of guardrail bypass that is difficult to prevent because it exploits the legitimate use case of design tools, requiring AI systems to infer intent rather than evaluate content alone
3. Espionage infrastructure — The use of AI-generated military IDs in espionage operations demonstrates how AI tools can lower the barriers to creating convincing identity documents, a capability previously requiring specialized forgery expertise
4. Dual-use ambiguity — The incident highlights the fundamental challenge of AI content policies where the same output (a military ID image) can be either a legitimate design sample or a forgery tool depending on user intent