CMMC 2.0 is how the US Department of Defense verifies that a supplier protects the information it handles. Level 1 covers federal contract information with 17 basic practices and an annual self-assessment. Level 2 covers controlled unclassified information with 110 practices taken from NIST SP 800-171. It is not a security framework you adopt for its own sake. It is a condition of holding the contract.
That changes the economics of getting ready. A SOC 2 gap costs you a slower sales cycle. A CMMC gap costs you eligibility. And because the requirement flows down the supply chain, plenty of companies discover the obligation when a prime contractor sends a flowdown clause rather than when they bid on anything themselves.
What each level asks for
#| Level 1 | Level 2 | |
|---|---|---|
| Protects | Federal contract information | Controlled unclassified information |
| Practices | 17 basic safeguarding practices | 110 practices, drawn from NIST SP 800-171 |
| How it is assessed | Annual self-assessment with a senior official affirmation | Self-assessment or an assessment by an accredited third-party organization, on a three-year cycle, depending on the contract |
| What infrastructure covers | Most of it, since the practices are elementary technical hygiene | A large share of the technical families, and none of the human ones |
| Where the work actually goes | Turning things on and proving they stayed on | Scoping, the system security plan, and closing the gaps the plan admits |
Level 1 is achievable by a small team in a quarter if the foundation is sound. Level 2 is a program. The difference is not the count of practices; it is that Level 2 requires a written system security plan describing how each requirement is met for a defined boundary, plus a plan of action and milestones for everything not yet met, plus a score submitted under the department’s assessment methodology, which weights requirements by impact so that some gaps cost far more than others.
Which practices land on infrastructure?
#NIST SP 800-171 groups its requirements into 14 families. They do not split evenly between engineering and organization, and knowing which is which is the fastest way to plan the work.
| Family | Mostly infrastructure | What it looks like in a cloud foundation |
|---|---|---|
| Access control | Yes | Single sign-on as the only human path, no identity users or long-lived keys, permission sets scoped per environment, production read-only by default |
| Audit and accountability | Yes | Organization-wide trail with log file validation, a separate archive account, configuration recording in every account, alarms on privileged actions |
| Configuration management | Yes | Everything in version control, a plan on every pull request, drift detection that opens an issue rather than silently reverting |
| Identification and authentication | Yes | MFA enforced above the account by policy, federated deploy identities, dormant access flagged automatically |
| System and communications protection | Yes | Network isolation domains, centralized egress, TLS everywhere, customer-managed keys with rotation |
| System and information integrity | Yes | Threat detection in every account, image scanning, centralized findings |
| Media protection | Partly | Encryption defaults and key management help; handling procedures for physical media do not exist in a cloud account |
| Maintenance, physical protection | No | Your offices, your equipment, your procedures |
| Awareness and training, personnel security | No | People, not systems |
| Incident response | Partly | Detection, alert routing and the record are infrastructure; the plan, the roles and the exercise are yours |
| Risk assessment, security assessment | Partly | Continuous control findings feed both; the assessment itself is a process you run |
Roughly: the foundation answers the families about machines, and you answer the families about people. Any vendor that claims otherwise has not read the families about people.
What the two conformance packs do
#The AWS Enterprise Baseline ships 29 conformance pack templates, two of which are CMMC-shaped: a Level 1 pack of 70 rules and a Level 2 pack of 130 rules. Both are off by default, each behind a single flag, with the running cost stated next to the switch. Targeting Level 1 means enabling the Level 1 pack; targeting Level 2 means enabling both, because the Level 2 practices build on the Level 1 ones rather than replacing them.
Security Hub has no CMMC standard, so the pairing that works is the packs for practice-level scoring plus AWS Foundational Security Best Practices for control-level findings across more than 300 controls. The Level 2 template also includes a small number of process checks that exist to record manual attestations — useful, and a reminder that even the automated view of Level 2 has a human-signed component.
The important thing about both packs is what they are made of. They are AWS Config rules evaluating controls that are already on: the organization trail, configuration recording in all 15 accounts, threat detection everywhere, the service control policies that deny identity users and enforce MFA above the account, the isolation domains between production and non-production. The packs do not install controls. They score them, per account, continuously, and give you something to paste into a system security plan that is true on the day the assessor reads it.
Scoping is the decision that sets the price
#The most expensive CMMC mistake is not a missing control. It is never deciding where controlled unclassified information is allowed to live, and therefore having to assess everything.
CMMC scoping distinguishes between assets that handle that information, assets that provide security protection for those, assets that are managed under your risk process but are not intended to handle it, and specialized assets. The categories matter because they determine how each asset is assessed and documented. What they have in common is that every one of them requires you to know where the data is — which is a question about boundaries, not about tooling.
An enclave is not a trick to avoid the requirements. It is a decision to apply them intensively to a small, defined set of systems instead of thinly to everything. The account boundary is the natural edge, for the same reason it is the natural edge for PCI scope reduction: it is enforced by the provider, it holds when something inside it fails, and it turns “what is in scope” into a list rather than a survey.
The discipline the enclave demands is about flow, not storage. Controlled unclassified information arrives in email, in tickets, in a file someone drops into a shared drive, in a screenshot pasted into a chat. A perfect enclave with a leaky intake is a documentation exercise. That part is process work, and it belongs to the organization.
What the evidence looks like
#For a Level 2 assessment, the artifacts an assessor works from are yours: the system security plan, the plan of action and milestones, and the record of the practices operating. Infrastructure supplies the third of those, continuously, which is what makes the first two writable without a research project.
| Assessor question | What answers it |
|---|---|
| How is each practice implemented in the boundary? | The system security plan, written by you, referencing the controls and their configuration |
| Is it implemented now? | Conformance pack scores per account and control findings per resource |
| Was it implemented throughout? | Configuration history, the organization trail, and the pull request record of every change |
| Who has access and who had it last quarter? | Single sign-on assignments, permission set definitions, and dormant access findings |
| What happens when something fails? | Threat detection findings, alert routing, and the record of what was done about each one |
| What is not yet met, and when will it be? | The plan of action and milestones, fed by the failing findings rather than by memory |
The evidence side of this works the same way for every framework in the list, which is the argument for building the evidence pipeline once rather than per audit.
Where this gets genuinely hard
#Three situations where the honest answer is that infrastructure design is no longer the main problem.
External cloud service providers. Defense contract clauses set expectations for cloud services that process covered defense information, including a requirement that an external provider meet a security baseline equivalent to FedRAMP Moderate, along with specific incident reporting and media preservation duties flowed down by contract. That is a question about the provider and the service, answered by authorization packages and contract terms, not by the way you have arranged your accounts.
Export-controlled work. If the data is subject to export control, access restrictions by nationality and residency come into play, and those are typically answered with a government cloud region and a personnel process. Government-region deployment is not something we offer, and no amount of commercial-region architecture substitutes for it.
The assessment itself. Only an accredited organization can perform a third-party assessment. We are not one, we do not employ one, and we have no certification program of any kind. Our mappings — including the documented mappings to NIST SP 800-171 Rev 2 and FedRAMP Moderate — are a lookup aid that tells you where to point an assessor. They are not an authorization and they are not an opinion.
A sequence that works
#- Find out what you actually handle. Federal contract information and controlled unclassified information are different obligations, and plenty of suppliers assume the higher one applies when it does not, or the lower one when it does not.
- Draw the boundary before buying anything. Decide the enclave, decide what counts as a security protection asset, and write both down.
- Turn on the recorders so the history exists. Every question in an assessment is about a period, and the period starts when the recording does.
- Enable the pack for your target level at least two weeks before anyone reads the score, and use the failing rules as the first draft of the plan of action and milestones.
- Close the gaps in score order, because the assessment methodology does not weight every requirement equally.
- Write the system security plan against the real configuration, not against the intent. If the plan and the environment disagree, the environment wins, and the assessor will find it.
If you want to know how much of the technical half you already have before committing to a date, the BuiltForProd Assessment measures an existing environment against the BuiltForProd Standard, and the full list of frameworks with documented mappings is on the compliance readiness page. The same infrastructure that answers a defense flowdown clause is the infrastructure that answers an enterprise security questionnaire, which is usually the next thing that lands.