AI Agents in the Defense Supply Chain: Running a Digital Workforce Under CMMC :

Defense contractors want the same thing every operations leader wants from AI agents: less time gathering and drafting, more time deciding. Around controlled unclassified information the rules are strict, and in 2026 they also moved. This guide explains where CMMC stands, what applies today, and how a digital workforce fits.

By Tim Bryant, LYFYEPublished Updated
Listen to this article9 min

Narrated with an AI voice from the text on this page.

Key takeaways
  • Answer first: the Department suspended CMMC Phase 2 on July 13, 2026, pending a program review. Phase 1 self-assessment requirements and your DFARS 252.204-7012 obligations still apply.
  • Any agent that stores, processes, or sends CUI is inside your assessment scope, and so are the model services and tools it uses.
  • The cheapest control is keeping CUI away from agents that do not need it.
  • Treat the suspension as time to prepare, not as a reprieve. The underlying NIST SP 800-171 requirements did not change.
How LYFYE builds

A digital workforce of specialized AI agents, built for your company and its departments. Each agent has a named job, its own access, approval steps for the decisions that matter, and a log a person can read. Your people stay in charge.

Where CMMC stands in September 2026

The CMMC program rule, 32 CFR part 170, took effect on December 16, 2024. The acquisition rule that puts CMMC into DoD contracts took effect on November 10, 2025, which began Phase 1 of a four-phase rollout. Phase 1 relies on self-assessments. Phase 2, which would have begun requiring third-party Level 2 certifications in applicable solicitations on November 10, 2026, was suspended with immediate effect on July 13, 2026, while the Department reviews the program. In September 2026 a DFARS class deviation told contracting officers to limit CMMC requirements to the self-assessment levels for now. No restart date has been published. Check the DoD CIO's CMMC page before you plan against any date, because this is still moving.

What still applies today

The suspension paused a rollout. It did not remove the security requirements under it.

  • DFARS 252.204-7012 still requires you to protect covered defense information using NIST SP 800-171 and to report cyber incidents to DoD within 72 hours of discovery.
  • Where a contract calls for it, you still post your self-assessment score in the Supplier Performance Risk System and affirm compliance.
  • Level 1 still covers the 15 basic safeguarding requirements for federal contract information. Level 2 still maps to NIST SP 800-171 Revision 2, which has 110 requirements.
  • Where certifications are issued, Level 2 assessments recur every three years, with an affirmation each year.

Scope: agents are assets too

A digital workforce changes your CUI footprint. An agent that reads a drawing, summarizes a contract data requirement, or drafts a report from controlled data is an asset that processes CUI. So is the model service it calls, and so is any tool or connector it uses to reach the data. If an agent sends CUI to an outside cloud service, DFARS 252.204-7012 expects that service to meet security requirements equivalent to the FedRAMP Moderate baseline, which FedRAMP now labels Class C. The simplest design decision is also the most effective: split the work. Agents that handle scheduling, supplier questions, or public information never touch CUI. Agents that must touch it run inside the protected environment, on services approved for it.

Where human approval fits

Nothing in CMMC requires a person to approve every agent action. Around CUI, though, it is sensible to hold three kinds of action for a person: anything that sends CUI outside the protected environment, anything that changes a controlled record, and anything that grants or changes access. The agent prepares the action with its evidence. A named person approves, changes, or rejects it. The log records both, which also gives an assessor a clean way to see that access and changes are controlled.

A practical order of work

For a contractor that wants agents working now without creating a problem for its next assessment, this order keeps risk low.

  • Map the workflow and mark every step that touches CUI.
  • Start with a workflow that does not touch CUI at all, and prove the controls there.
  • For CUI workflows, confirm the model service and tools meet the 7012 requirements before building.
  • Give each agent its own identity, narrow permissions, and an entry in your system security plan.
  • Log every step, link approvals to actions, and include agent incidents in your incident response plan.
  • Update your self-assessment to reflect the new assets, and keep the evidence current.

Planning for the restart

It is tempting to read the suspension as permission to wait. That is a risk. Primes still flow down security requirements, incident reporting obligations did not pause, and a program review can end with changes that arrive on short notice. Building agents with identity, logging, and approvals in place from the start costs little more than building them without, and it avoids rework when third-party assessments resume in whatever form they take.

How LYFYE helps, and what we do not do

LYFYE designs and builds digital workforces for companies and their departments, including agents that work around controlled information, with each agent's identity, permissions, approval steps, and logs designed in and documented. LYFYE is not a certified third-party assessment organization, does not hold a CMMC certification, and does not certify anyone. Your assessor does the assessment and your company makes the affirmation. We build the agents so the evidence is there.

NIST SP 800-171 families that carry the most weight for agents

FamilyWhat it asksHow the agents meet it
Access Control (3.1)Limit access to authorized users and processes, with least privilegeEach agent has its own identity and reaches only the CUI its task needs
Audit and Accountability (3.3)Create, protect, and review records that trace actions to usersA trace per task, with the agent, the approver, and the result, stored where the agent cannot change it
Configuration Management (3.4)Baselines and controlled changesAgent instructions, tools, and model versions are versioned and changed under review
Identification and Authentication (3.5)Identify and authenticate users and processesAgents are identified separately from people. Approvers use multifactor sign-in.
Incident Response (3.6)Detect, handle, and report incidentsA stop switch and an agent incident runbook, tied to your reporting process
System and Communications Protection (3.13)Protect CUI at boundaries and in transitAgents and their model services stay inside the protected boundary
System and Information Integrity (3.14)Identify flaws and monitor for attacksOutside content is treated as data, and unusual agent behavior raises an alarm

Primary sources

Vendor capabilities change. Every claim above that depends on a provider's current behavior is checked against that provider's own documentation on the date shown.

Your next big move. Built with LYFYE.

Tell us about the work one of your departments does today. We will show you where a digital workforce fits, which decisions stay with your people, and what it would take to build, starting with one workflow.