North Korea: Microsoft measures, group by group, how deeply AI has entered attackers' tradecraft

On 29 October 2026, Microsoft Threat Intelligence published the report “Tracking AI adoption in North Korean threat actor tradecraft” on the Microsoft Security Blog. It examines eight groups that the company attributes to North Korea and names with 'Sleet' monikers, its taxonomy for these actors. Overall AI adoption is rated 'high' on a four-stage scale. In the most advanced stage, 'operationalized', where AI is built into routine operations, Microsoft places Jade Sleet, Coral Sleet and Jasper Sleet. Diamond Sleet, Onyx Sleet, Pearl Sleet and Sapphire Sleet sit in the 'accelerating' stage, Emerald Sleet in the 'expanding' one. No group remains in the initial 'emerging' stage. According to the company, the goals have not changed: generating revenue, gathering strategic intelligence and, in some cases, carrying out destructive operations.

The episodes described give the scale some substance. Microsoft reports that from 8 to 12 June Jade Sleet ran recruiting-themed attacks through three front companies, posing as recruiters for a non-existent AI-based blockchain intelligence firm. SBS's account, which does not name the group, says the fake offers targeted workers in AI, blockchain and cybersecurity. Between August and September, Onyx Sleet reportedly produced more than a hundred lure PDFs aimed at South Korean users, each with target-specific details, dozens at a time within minutes. Microsoft believes they were likely generated with AI. Since April 2026, Coral Sleet has reportedly used a workflow to plant trojans in private GitHub repositories, probably with the help of an AI-based development platform. The company also considers it likely that Sapphire Sleet used AI during the compromise of a piece of Rust software in August 2026. The newest finding concerns ARTEX, an open-source penetration testing framework built on agentic AI, which according to the report Onyx Sleet and Pearl Sleet use in their operations. SBS places the start of this use at the end of October, right before publication. For Emerald Sleet, Microsoft flags possible use of synthetic voice and video, but considers the operational impact uncertain.

Alongside the cases, the report lists recurring uses: reconnaissance on organisations, employees, wallets and cloud services; tailored phishing, recruiting and investment lures, with translation and tone adjustment; fake identities, CVs and answers for job interviews; code generation, debugging and refactoring, including with the Cursor platform; summarising stolen information. Back on 6 March 2026, a Microsoft report covered by CyberScoop had described Jasper Sleet's use of generative AI in fake remote IT worker schemes. Also in March, Microsoft had described AI as a force multiplier for attackers. The new document also draws a line, in its own words: “AI-enabled operations remain human-directed and have not become fully autonomous.” Among the recommendations are stricter candidate checks in hiring: identity and reference verification, searching for duplicate identities, confirming on-camera presence and real phone numbers. Then come anti-phishing defences, exposure management and AI governance.

All of this should be read for what it is. The group attributions and the judgements on AI use are Microsoft's assessments, often accompanied by a 'likely', and the individual episodes have not been independently verified. The report does not say which commercial models were used, apart from Cursor. What stands out is less spectacular than an autonomous attack, and more stubborn. AI has not invented new goals: it seems instead to make chasing the old ones faster and cheaper, with dozens of personalised lures within minutes or a job interview held by a candidate who does not exist. That is why the most telling recommendation is also the oldest: ask someone to turn on their camera.

— Olya