Working with Defense Industrial Base contractors every day, I hear a version of the same question:

What does adversary AI actually look like for a supplier trying to stay compliant and secure?

It usually does not look like autonomous malware running wild. It looks like familiar tradecraft moving much faster.

Microsoft Threat Intelligence’s research into North Korean threat actors shows generative and agentic AI becoming part of routine operations. The objective has not changed. These groups still rely on social engineering, malware delivery, identity deception, and legitimate tools. AI helps them perform those tasks with more speed, scale, and consistency.

That distinction matters for DIB organizations. The threat is not waiting for some future form of artificial intelligence. It is already using today’s tools to shorten the time between reconnaissance, targeting, initial access, and exploitation.

Social engineering at machine speed

Microsoft identified more than 100 target-specific PDF lures attributed to Onyx Sleet from August and September 2026. Metadata showed the actor creating dozens at a time, within minutes.

A tailored lure used to require more manual work: research the target, write the message, build the document, check the language, and repeat. Generative AI reduces that lead time. An attacker can create variations for different companies, roles, projects, or technical communities while the underlying campaign stays the same.

For defenders, this weakens one of the assumptions behind traditional awareness training. A phishing message no longer has to be generic, poorly written, or visibly suspicious. It can reference the right role, use credible industry language, and arrive with a document built for a narrow audience.

Personas and front companies can appear in days

Microsoft also observed Jade Sleet run job-themed targeting through three front companies in a single week. The campaigns used tailored domains, social profiles, recruiter personas, job descriptions, and industry knowledge to approach people working in cybersecurity, AI, blockchain, and recruiting.

That is especially relevant to the DIB, where trusted relationships often begin through recruiting, subcontracting, consulting, or technical collaboration. A convincing persona does not need to survive years of scrutiny. It may only need to earn one conversation, one download, or one set of credentials.

Identity verification now has to extend beyond the sign-in screen. Organizations need practical ways to validate outside recruiters, prospective partners, unusual meeting requests, code repositories, and files shared during hiring or business-development conversations.

Agentic tools are entering routine tradecraft

Onyx Sleet and Pearl Sleet have also used ARTEX, an open-source agentic penetration-testing framework, in operations. That does not mean the attack has become fully autonomous. Microsoft assesses that these operations remain human-directed.

The important change is operational tempo. Agentic tools can help automate reconnaissance, organize findings, test paths, and repeat tasks that previously consumed an operator’s time. The attacker still chooses the objective, but more of the work between decisions can happen automatically.

This compresses the kill chain. Defenders have less time to recognize a lure, investigate an endpoint, contain an identity, and determine whether sensitive data moved.

What this changes for a CMMC Level 2 program

For organizations building toward CMMC Level 2 and protecting CUI in Microsoft 365 GCC High or Azure Government, compliance controls still matter. But a control that exists on paper and operates too slowly will not stop an AI-accelerated campaign.

Three areas deserve attention:

  1. Protect email and collaboration traffic with active investigation. Defender for Office 365 can combine Safe Links, Safe Attachments, campaign analysis, and automated investigation and response. When attackers can regenerate targeted lures quickly, security teams need tools that can identify related messages, analyze URLs and attachments, and remove malicious content across mailboxes.

  2. Detect and contain behavior at the endpoint. Defender for Endpoint adds device-level visibility, behavioral detections, automated investigation, and remediation. A polished lure may still reach a user. The endpoint layer has to recognize what happens next and give defenders a way to isolate the device before access spreads.

  3. Protect the data even when an identity is compromised. Microsoft Purview sensitivity labels can classify CUI and export-controlled information, apply encryption, and restrict who can open or use the content. That protection can remain with a labeled file after it leaves its original SharePoint site or mailbox. Identity controls are essential, but the data needs its own protection boundary.

None of these products replaces the work required by NIST SP 800-171 or CMMC. They help make the controls operational. The goal is to connect identity, endpoint, email, and information protection so that one convincing message does not become unrestricted access to the enclave.

The baseline has shifted

Adversaries are not using AI to invent an entirely new attack chain. They are using it to remove friction from the one they already know.

For the DIB, that means faster lures, more convincing personas, broader reconnaissance, and less time to respond. Annual training, legacy filtering, and audit-time evidence collection are not enough on their own. The defensive program has to operate continuously and at a speed closer to the threat.

NIST SP 800-171 and CMMC are more than audit checkboxes. When their controls are implemented through a Zero Trust architecture, they create the identity, device, data, and monitoring boundaries that make an AI-accelerated attack harder to complete.

The attacker’s workflow is getting faster. Ours has to get harder to bypass.

Sources