Microsoft Threat Intelligence said on 29 October 2026 that North Korean threat actors have broadened their use of generative and agentic AI across reconnaissance, social engineering, software development and operational tasks. The main consequence for defenders is not autonomous attack activity, but a higher pace in workflows that already work for these actors.
The assessment says North Korean activity remains tied to existing objectives. Those include revenue generation for the regime, strategic intelligence collection and, in some cases, politically motivated disruptive or destructive activity. The common targets remain cryptocurrency and financial organisations, IT companies and software supply chains, government bodies and education-related organisations.
AI use supports established tradecraft
The assessment maps high-impact North Korean actors across stages of AI adoption. Jade Sleet, Coral Sleet and Jasper Sleet are described as having embedded AI into routine operational workflows. Diamond Sleet, Onyx Sleet, Pearl Sleet and Sapphire Sleet are placed in a rapid adoption stage, while Emerald Sleet is described as expanding AI use in selected activities.
No North Korean threat actor is currently placed in the limited or experimental stage. That does not mean the activity has become fully automated. The assessment states that AI-enabled operations remain human-directed.
The practical change is in throughput and preparation. AI helps operators research targets, draft more relevant lures, maintain fake identities, work across languages, generate or refactor code and process collected information. Those uses reduce the time needed to prepare campaigns and make repeated modification easier after defenders disrupt a tactic.
Reconnaissance and lures
Jade Sleet activity between 8 June and 12 June 2026 shows the pattern. The actor conducted job-themed targeting against people in organisations connected to cybersecurity, AI, blockchain and talent recruitment. It used tailored domains, social media profiles and sector knowledge to pose as recruiters for a spoofed AI blockchain intelligence company.
During the same period, Jade Sleet used two additional front companies with comparable infrastructure and domain knowledge. The assessment links the pace and breadth of that activity to generative AI use in target research, industry preparation and outreach.
Onyx Sleet provides another example. More than 100 PDF lure files attributed to the actor were created in August and September 2026 with target-specific details. Metadata showed dozens of files created within minutes, which the assessment says was highly likely the output of an AI platform.
That matters for mail and collaboration security teams because lure regeneration can be fast. Blocking one set of messages or URLs may not be enough if the actor can produce modified files and related content quickly. The defensive work then moves beyond single indicators into behaviour, sender patterns, identity controls and user reporting.
Fake identities and employment operations
Several North Korean actors use employment or interview themes to reach targets. Sapphire Sleet, Jade Sleet and Coral Sleet use job-themed social engineering against users and organisations. Jasper Sleet uses fraudulent identities to obtain legitimate remote employment.
AI supports these operations by generating resumes, biographies, interview responses and job-related communication. It also helps maintain multiple worker personas and adjust language or tone during interactions. The assessment also identifies possible synthetic media generation by Emerald Sleet, focused on AI-created voice and video, although the full operational impact is not stated.
For hiring and supplier processes, the guidance is more operational than technical. It recommends scrutiny of staffing-company workers, checks for repeated or highly similar professional profiles, consistency checks across names, addresses and dates, and video participation in multiple calls. It also recommends confirming that a candidate has a real digital footprint, including a non-VOIP phone number, a residential address and social media accounts.
Tooling, code and agentic frameworks
AI-assisted development tools are also part of the activity. The assessment identifies evidence that Jade Sleet, Coral Sleet, Pearl Sleet and Sapphire Sleet used AI-enabled development platforms such as Cursor to generate, convert, debug and refactor code. The same tooling can help with scripts, commands, persistence mechanisms, obfuscation and troubleshooting.
In the August 2026 Rust software compromise attributed to Sapphire Sleet, the malware showed a marked improvement compared with other attributed campaigns. It used dynamic .NET compilation during post-exploitation, most likely for defence evasion. Since April 2026, Coral Sleet has also operated a workflow to trojanise private GitHub repositories and integrate them into its existing malware chain.
Agentic tools are now appearing in the same activity set. Since late October 2026, Onyx Sleet and Pearl Sleet have been observed using ARTEX, an open-source agentic AI penetration testing framework, in operations. For security teams, that places open-source agent tooling inside the threat model rather than only inside internal red-team or testing workflows.
Defensive controls
The guidance groups defensive work around exposure management, social engineering protection, employee vetting and malware defence. It recommends Microsoft Defender Vulnerability Management, Defender External Attack Surface Management, secure network boundaries for web-facing systems, automated patching where possible and Azure Update Manager for distributed infrastructure.
For phishing and social engineering, the recommended controls include Exchange Online Protection, Microsoft Defender for Office 365, attack simulation training, Zero-hour auto purge, Threat Explorer, Safe Links and Safe Attachments. It also recommends blocking device code flow wherever possible and allowing it only where necessary through Conditional Access policies.
For malware, the guidance recommends cloud-delivered protection in Microsoft Defender Antivirus or an equivalent product, current definitions, real-time protection, endpoint detection and response in block mode, full automated investigation and remediation, automatic attack disruption in Microsoft Defender XDR and network protection in Microsoft Defender for Endpoint. The listed attack surface reduction rules include blocking executable content from email client and webmail, blocking executable files unless they meet prevalence, age or trusted-list criteria, blocking obfuscated scripts and using advanced protection against ransomware.
The post also points to Microsoft Security Dashboard for AI as a way to view AI security posture across Microsoft Defender, Microsoft Entra and Microsoft Purview. Microsoft Agent 365 is described as a control plane for managing, governing and securing AI agents and workflows, including Microsoft agents, agents from Adobe and Databricks, and open-source agents published on GitHub.
The risk that AI-generated lures and agentic testing frameworks create for Defender, Sentinel and Purview operations is the practical question ECS's Security topic deals with, alongside Security Copilot, Zero Trust and DevSecOps. Security sessions and workshops give attendees room to compare detection, response and governance work with people running these controls in production. Bring the threat models, tenant settings and incident response questions that are hard to settle from a blog post alone.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.