Microsoft rates AI adoption across eight North Korean threat groups on a four-stage scale, placing three at 'operationalized' while judging operations still human-directed

Uses from fake recruitment to mass-produced lures

The uses the report lists include reconnaissance and bespoke lures, fabricated personas and resumes (Jasper Sleet), interview-themed social engineering (Sapphire, Jade and Coral Sleet) and malware development in AI-assisted coding environments (Jade, Coral, Pearl and Sapphire Sleet). More than 100 PDF lure files attributed to Onyx Sleet were created in rapid bursts in August and September 2026, with the pace of document production offered as an indicator of AI use.

Jade Sleet ran job-themed targeting through three front companies in a single week, 8–12 June 2026, and Sapphire Sleet likely used AI in the August compromise of Rust software. Coral Sleet has been trojanizing private GitHub repositories since April, and Emerald Sleet is flagged as a possible user of synthetic voice and video.

Arrival at agentic penetration-testing tools

The notable passage is that Onyx Sleet and Pearl Sleet began using the open-source agentic penetration-testing framework ARTEX in late October 2026. Microsoft's 21 October report described ARTEX in a supply-chain compromise by the China-linked Storm-0940, and TrendAI's 27 October report found it in attack infrastructure aimed at Japan. The same agentic tool spreading across state actors within weeks has therefore now been shown independently by several observers.

What the 'human-directed' verdict covers

Microsoft assesses that AI is increasing the speed and efficiency of attacks while humans still direct operations and no fully autonomous attack has occurred. The verdict gives a baseline for the current stage of attacker automation against which defenders' early access to frontier models and pilots of autonomous response can be measured. The recommendations combine conventional controls, including vulnerability management, EDR block mode, attack-surface-reduction rules, vetting of staffing firms and blocking the device-code flow, with visibility into AI agents.