Microsoft Threat Intelligence has published an assessment of how North Korean state-linked groups are using artificial intelligence, and the headline finding is a classification rather than an incident: adoption across the actor set is high, and for three groups it is now embedded in routine work rather than bolted on.
Microsoft sorts the groups along a four-stage ladder — emerging, expanding, accelerating and operationalised. Jade Sleet, Coral Sleet and Jasper Sleet sit at the top, with AI written into their ordinary workflows. Diamond, Onyx, Pearl and Sapphire Sleet sit below them, adopting AI rapidly across several functions, and Emerald Sleet uses it in selected activities.
What the evidence looks like
The assessment rests on operations Microsoft says it observed. Jade Sleet ran campaigns through three front companies in a single week — job-themed outreach that would previously have taken a team far longer to stage at that quality. Onyx Sleet is attributed with more than 100 target-specific PDF lures created in August and September 2026, dozens of them produced within minutes, a tempo Microsoft says was highly likely enabled by an AI platform.
Coral Sleet is the more consequential finding. Since April 2026 the group has run a workflow to trojanise private GitHub repositories and fold them into its existing malware chain, a capability Microsoft assesses was highly likely built on an AI-enabled development platform. Sapphire Sleet likely used AI during an August 2026 compromise involving Rust software.
AI-assisted development tools are named in the report. Microsoft says platforms such as Cursor have likely let Jade, Coral, Pearl and Sapphire Sleet generate, convert, debug and refactor code, build scripts, commands and persistence mechanisms, and troubleshoot their own tooling — the same productivity gain defenders chase, applied to intrusion.
The agentic turn
Since late October 2026, Microsoft says it has observed Onyx Sleet and Pearl Sleet using ARTEX, an open-source agentic penetration-testing framework, in operations. ARTEX is not bespoke malware; it is a public tool that runs agents against a target, accumulates what they find into a graph and reports back. Microsoft linked it to a nation-state campaign against ConnectWise earlier in the month, and its appearance in North Korean operations is a sign that commodity agentic tooling is being absorbed into state tradecraft at the same speed it is absorbed into enterprise security testing.
Human hands, faster
Microsoft is careful about one thing: the operations remain human-directed. AI has not made these groups autonomous. What it has done is compress the time between intent and artefact — the front companies, the lures, the trojanised repositories — and widen the number of targets a single operator can service.
That distinction is what makes the report useful to defenders. "Operationalised" is a statement about workflow, not about a new exploit class: detection content written for the malware these groups already deploy still applies, which is why Microsoft points customers at existing Defender detections rather than announcing a new threat category.
What it changes
The report does not describe a novel attack so much as a change in the cost of the old ones. Tailored lures used to be evidence of a well-resourced actor; they are becoming evidence of a well-configured one. For organisations in the sectors these groups target — defence, crypto, recruitment, and increasingly AI itself — the practical implication is that socially engineered contact should be treated as routine rather than exceptional, and that a private repository is a target surface, not a back office.
Sources: Microsoft Threat Intelligence — Tracking AI adoption in North Korean threat actor tradecraft; Microsoft — Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise.