AI Is Now Standard Kit: Microsoft Puts Three North Korean Groups at Full Operational AI Adoption

On 29 October 2026 Microsoft Threat Intelligence published its most structured accounting yet of how North Korean threat actors use AI — and the headline is maturity, not experimentation. Microsoft assesses overall DPRK adoption of AI as high, places three groups at full operationalization (AI embedded across entire operational workflows), and assesses that no tracked DPRK actor remains at the merely "emerging" stage. The operations are still human-directed — Redmond is explicit that AI-enabled activity has not become fully autonomous — but AI now accelerates and scales the established playbook: revenue generation, strategic intelligence collection, and in some cases disruptive objectives, built on social engineering, malware delivery, and abuse of legitimate tools against cryptocurrency, IT and supply-chain, government, and education targets.

The reason this belongs on an agent-security site rather than a geopolitics one: the report documents DPRK operators using agentic AI tooling as ordinary tradecraft — including the same open-source agent frameworks defenders track in the lab. Onyx Sleet and Pearl Sleet are leveraging the open-source agentic AI penetration-testing framework ARTEX to conduct operations. Readers of this site's VAPOR PANDA ARTEX espionage coverage and the Storm-0940/Coral Sleet ConnectWise attribution will recognize the framework; Microsoft now ties it directly to North Korean operational use, which is a second, independent confirmation that ARTEX has crossed from research artifact into state-actor kit.

The four stages, and who sits where

Microsoft maps high-impact DPRK actors across four adoption stages. Operationalized — AI embedded in routine workflows — covers Jade Sleet, Coral Sleet, and Jasper Sleet, whose operations frequently center on identity-based attacks built on deceptive social engineering, now accelerated with AI-tailored lures and personas. Accelerating — rapid adoption across multiple operational functions — covers Diamond Sleet, Onyx Sleet, Pearl Sleet, and Sapphire Sleet. Expanding — AI in selected activities but not embedded across workflows — covers Emerald Sleet. The emerging bucket is empty: Microsoft has not assessed any North Korean actor as limited/experimental. That empty bottom rung is itself the finding — the floor has risen across the whole set.

The case studies make the taxonomy concrete. Jade Sleet ran campaigns through three front companies in a single week (8–12 June 2026): job-themed targeting against cybersecurity, AI, blockchain, and recruitment personnel, posing as recruiters for a spoofed AI-powered blockchain intelligence company with tailored domains, social profiles, and industry-specific knowledge — a pace and breadth Microsoft attributes to generative AI accelerating targeting knowledge, reconnaissance tailoring, and outreach scaling. Onyx Sleet generated dozens of tailored lures in minutes. Coral Sleet rapidly built a new capability to trojanize private source-code repositories. Diamond Sleet used AI to sharpen targeting workflows including persona development and front-company creation — fake communications-technology firms with supporting websites and personas aimed at telecoms — and since August 2026 has returned to the oil and gas sector for the first time since 2023 while expanding into communications and internet-policy organizations, shifts occurring alongside its increased generative-AI use for reconnaissance and development. Jasper Sleet continues building fake digital personas, with suspected agentic-AI-assisted resume workflows and repeatable tooling that scale employment-themed infiltration — the remote-IT-worker pipeline — with improving speed, scale, and relevance.

What AI actually touches in the kill chain

The report is useful because it itemizes AI use by function rather than gesturing at "AI-powered attacks." In targeting and reconnaissance: researching organizations and employees to craft bespoke lures and job applications; researching wallets, exchanges, and DeFi workflows to accelerate domain knowledge for crypto targeting; probing vulnerabilities, APIs, security products, and cloud services — with, Microsoft notes, a demonstrated change in the speed and sophistication of activity against specific technologies; and researching roles and requirements to support persona alignment and rapid spear-phishing. In social engineering and identity deception — the core of Sapphire, Jade, and Coral Sleet's employment- and interview-themed operations and Jasper Sleet's fraudulent-employment play: tailored phishing, recruiter, and investment lures; job-themed payload-delivery content and malicious repository lures; translation and tone adjustment for foreign-language impersonation; resumes, biographies, and interview responses; coding assessments and multi-persona management; and altered profile images, video, voice, and backgrounds to sustain recruiter, developer, and investor identities across interactions.

Note the through-line for defenders: AI is a force multiplier on existing tradecraft, not a new capability class. The actors did not change motivations or targeting priorities — sanctions evasion, revenue, intelligence — they compressed the expensive parts (persona building, lure tailoring, domain research, outreach volume) that used to bound campaign scale. Controls that assumed human-speed personalization — "a recruiter who knows our stack this well must be legitimate," interview processes that verify one persona at a time — degrade first.

What to do

• Re-price employment-themed social engineering. Jade, Coral, Sapphire, and Jasper Sleet all run hiring-pipeline lures; AI now generates the tailored domains, profiles, resumes, and interview performance. Verify recruiters and front companies through independent channels, and treat coding assessments and repository invitations from hiring contacts as payload-delivery risks.
• Watch the ARTEX overlap. With Onyx and Pearl Sleet using the ARTEX agentic pentest framework operationally, ARTEX-associated infrastructure, tooling markers, and TTPs belong in DPRK threat models — cross-reference this site's VAPOR PANDA ARTEX briefing and the Storm-0940 ConnectWise machine-speed attribution rather than treating each as an isolated report.
• Apply Microsoft's published detections. The report ships mitigation and protection guidance with Microsoft Defender detections — ingest those, and hunt for the AI-compressed precursors: bursts of newly registered tailored domains, high-volume personalized outreach, and fake-company infrastructure stood up in days rather than weeks.
• Do not confuse "human-directed" with "slow." Microsoft's judgment that operations remain human-directed is not comfort: dozens of tailored lures in minutes and three front companies in a week are human-directed outputs at machine speed. Rate-limit assumptions in detection engineering should be revisited.

Sources:

• Microsoft Threat Intelligence — "Tracking AI adoption in North Korean threat actor tradecraft" (29 October 2026; adoption stages, Sleet assessments, case studies, mitigations and Defender detections; read first-hand)
• Chosun Biz — "North Korean hackers wield AI to scale phishing and malware against South Korea" (30 October 2026; press coverage of the report)
• ChainCatcher — Microsoft Threat Intelligence report summary via Yonhap (29–30 October 2026; eight tracked groups, crypto-sector targeting)