The government just wrote its first contract clause for what happens to your data inside an AI tool. It took effect today.

A General Services Administration clause effective October 19 governs government data that goes into or comes out of a large language model: a 72-hour reporting clock, access that must be logged and time-bound, deletion of fine-tuned model weights at the end of the order, and precedence over whatever the vendor’s own terms of service say. It is named after the basic safeguarding clause — and the Department of War is already under a statutory direction to build its own version into CMMC.

The clause is GSAR 552.239-7001, and its title is the part to notice: Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems. For a decade the government’s floor for contractor information systems has been the fifteen “basic safeguarding” requirements of FAR 52.204-21 — renumbered 52.240-93 under this year’s acquisition overhaul, and the same fifteen that sit behind a CMMC Level 1 self-assessment. GSA chose those two words deliberately. This is not a policy statement about artificial intelligence. It is a floor, written as a contract clause.

It arrives the same way the CMMC pause did — as a class deviation rather than a rule. GSA issued it under its acquisition-overhaul deviation authority, with formal rulemaking and a public comment period to follow later. Contracting officers had discretion to use it before the effective date, and GSA plans to fold it into the Multiple Award Schedule solicitation at the November refresh.

  • Oct 6, 2026
    GSA posts advance notice of the November schedule refresh, including the new acquisition-regulation language and the clause. Comments open for ten business days.
  • Oct 13, 2026
    Final clause language released. Practitioner alerts begin walking the text the same week.
  • Oct 19, 2026 — today
    552.239-7001 takes effect as a class deviation, functioning as GSA policy until it is rescinded or codified.
  • Oct 26, 2026
    GSA webinar on the refresh.
  • Nov 6, 2026
    Schedule Refresh #34 planned, adding the clause to the solicitation. Trade reporting puts the acceptance window for the accompanying mass modification at 90 days.

Who it reaches: contractors offering a large language model product or service to the government — a chatbot, an agentic system, an LLM-enabled productivity tool — where government data is submitted to or produced by the model. It is self-deleting where the model lives only inside a contractor’s own back-office systems and is never delivered to the government, and where LLM functionality is incidental to a product that is not an AI product. It flows down to subcontractors performing AI design, development, deployment, or operation and monitoring tied to government data, with the prime carrying the best-efforts and due-diligence obligation for the flow-down.

What it requires, according to the practitioner analyses published in the week since the final text appeared: access to government data only on a demonstrated operational need, for a specific purpose, time-bound, logged, and producible on request. Incident reporting on a 72-hour clock, narrowed to incidents affecting a model used in contract performance that may affect government data — with FedRAMP or CISA reporting capable of satisfying it in some cases. A bar on training models with government data. Rights in improvements or enhancements built from government data assigned to the government. At the end of the contract or order, deletion of government data and custom developments, including fine-tuned model weights and embeddings, unless the contracting officer directs otherwise in writing. And the clause takes precedence over conflicting contractor terms — which is the quietly consequential part, because it means a click-through terms-of-service page no longer governs what happens to the government’s data.

The Department of War’s version already exists — as a direction, not a clause: Section 1513 of the FY2026 National Defense Authorization Act directs the Department to develop and implement a framework for the cybersecurity and physical security of the AI and machine learning it acquires, aimed at data poisoning, adversarial tampering and unintentional data exposure. Three details matter for this audience. The framework must be built as an extension or augmentation of existing Department cybersecurity frameworks, CMMC named among them. It is to be informed by the NIST 800 series. And it is to be imposed on contractors through a DFARS amendment, after a cost-benefit analysis weighing security against the cost of slowing AI adoption. “Covered” AI and machine learning includes source code and model weights. There is no statutory deadline — a status report to Congress was due in June 2026 — so treat the timing as unknown and the direction as settled.

Three features of the GSA clause are worth holding up against the rules you already live under. A 72-hour reporting clock. Access that is purpose-bound, time-bound and logged. And a refusal to let a commercial vendor’s terms override the government’s. Those are the same three instincts behind DFARS 252.204-7012 and the 110 requirements of NIST SP 800-171. The only thing that has changed is the subject — not a network this time, a model.

So what: you do not need a Department of War AI rule to have an AI problem. If anyone in your organization has pasted a drawing, a specification, a test result or a requirements document into a commercial chatbot, controlled data left your boundary already, and no clause had to publish for that to be true. This week, write down three columns: which AI tools touch controlled data, who approved each one, and what that vendor’s terms say about retention and training. It is a short list and an afternoon’s work — and it is the first thing every rule now in motion is going to ask you for.

The Department has already written down what it means by “the basics.” One of the twenty items has no counterpart in the 110.

Last week’s brief covered the launch of the Department of War CIO’s “Brilliant at the Basics” campaign on October 1 and the four weekly themes carrying it through the month. The substance underneath the campaign is older, and considerably more useful: two one-page sheets listing ten information-technology practices and ten operational-technology practices, published over the summer alongside the July 13 suspension. The sheets carry no visible revision date — the IT sheet is marked version 1, the OT sheet version 2. This is the full set.

#The IT tenThe OT ten
1Phishing-resistant multi-factor authenticationIdentity and access control
2Comprehensive asset inventory managementValidated asset inventory
3Strategic technical debt reductionStrict network segmentation
4Flexible technology stackOT-specific incident response and recovery plan
5Logical segmentation to limit adversary lateral movementManage known vulnerabilities
6Risk-based vulnerability managementRemote access pathways
7Integrate security early in the development lifecycleContinuous monitoring
8Secure AI adoption and data protectionSystem resiliency
9Resilient backup and disaster recovery architectureSupply chain security
10Continuous technical workforce readinessReview processes

Read the sheets themselves and the framing is unusually candid. They carry no assessment method, no scoring rubric, no evidence standard and no certification path, and the Department’s own disclaimer states the material is educational, that it accepts no liability for losses arising from its use, and that following it does not guarantee immunity from threats. There is no such thing as being compliant with this list, and nobody will ever ask you to prove you are.

What is worth knowing is how much of your actual obligation these twenty practices cover. No official crosswalk between them and the 110 requirements of NIST SP 800-171 has been published. Two independent analyses have each built their own and reached figures in the same neighborhood: one maps the IT ten onto nineteen of the 110, worth at most seventy-one of the weighted points; the other counts twenty-four unique requirements spread across eight of the fourteen families. Call it roughly a fifth, with the caveat that both numbers are their authors’ editorial judgment rather than anything authoritative. Either way the conclusion holds in the same direction — the floor the Department describes sits well below the bar you already signed for.

The omissions are the telling part, and they are consistent: governance and documentation. No system security plan. No plan of action and milestones. No audit logging requirement. No personnel screening records. Those are precisely the artifacts that make a score defensible when somebody asks to see it, and precisely what the enforcement cases on the public record have turned on. In one place the list is genuinely harder than the baseline — it asks for phishing-resistant multi-factor authentication, where the Revision 2 requirement asks only for multi-factor.

Item eight, and why it connects to the story above: the IT sheet’s eighth practice is secure AI adoption and data protection: set policies and technical controls for AI tools, prohibit entering sensitive Department data into public commercial AI systems, and favor approved enterprise environments instead. One of the two published crosswalks flags it as the single item on the list with no Revision 2 analogue at all. So the Department has put AI on its own list of fundamentals, while the standard your SPRS score is measured against does not mention it. That gap is exactly where the clause in the story above just landed — on the civilian side, first.
So what: take the twenty practices and mark each one evidenced, believed or unknown. One afternoon, no tooling, and the “unknown” column is your next quarter’s work plan. What you should not do is budget against this list — it is a fifth of the requirement set and none of the paperwork. And give item eight an owner, by name, this week: it is the one line on the Department’s own floor that your 110-requirement program is not already trying to cover.

While everyone watched Level 2, the standard above it was quietly rewritten

On May 13, 2026, NIST finalized Revision 3 of SP 800-172 and its companion assessment publication, SP 800-172A. These are the enhanced security requirements for controlled unclassified information on non-federal systems — the layer above the 110, applied when the information is associated with a high value asset or a critical program, and written against advanced persistent threats rather than opportunistic ones. It went largely unremarked in the defense industrial base, because the attention that month was fixed on the program below it.

The scale of the rewrite is not incremental. NIST added eighty new requirements, withdrew twelve, and significantly changed twelve more, reorganized across seventeen families around a three-part strategy: penetration-resistant architecture, damage-limiting operations, and cyber resiliency. Much of the new material is acquisition and supply chain risk management; the rest lands on access control, network segmentation, asset management and threat detection. A number of requirements now carry organization-defined parameters, meaning the agency or the organization picks the actual value — a tool, a mechanism, a time period — rather than reading it off the page.

CMMC Level 3 does not run on this yet. It still requires a Level 2 certification, twenty-four selected requirements drawn from the February 2021 version, and a government-conducted assessment, and the Department has said it will update its security requirements through later rulemaking. So if your reaction is “Level 3 isn’t me,” you are right about Level 3 and wrong about the exposure. The live mechanism is agency discretion: practitioner guidance on the final publication is explicit that agencies may impose portions of SP 800-172 on contracts, grants or other agreements involving particularly sensitive data without waiting for CMMC to catch up. One civilian agency was already citing selected draft 800-172 requirements in its own CUI guide, as this brief covered last week. The draft is now final.

So what: find out whether any program you support is designated a high value asset or a critical program. Your contracting officer knows; you very likely do not, because nothing in your contract file announces it in those words. That single question decides whether SP 800-172 Revision 3 is background reading or a requirement sitting in wait inside a future solicitation — and the eighty new requirements are not something you absorb in a bid cycle.

What to do this week

The theme this issue: the newest rule is about a tool your people are already using.

  • Inventory the AI tools that touch controlled data. Which tool, who approved it, what the vendor’s terms say about retention and training, and whether anyone has pasted a drawing or a specification into it. Put one name against the list — someone with authority to approve or refuse a tool.
  • If you hold a GSA schedule or any civilian contract, read 552.239-7001 before you accept the November mass modification. The clause is effective now; reporting puts the acceptance window at 90 days from the refresh.
  • Ask your contracting officer whether any program you support is a high value asset or a critical program. That answer decides whether SP 800-172 Revision 3 is reading or work.

Quick hits

  • Oct 12, 2026
    The Pentagon and the SBA signed the financing half of the industrial-base push.
    A memorandum of agreement — SBA release 26-104 — builds on the Smaller War Plants Commission the two agencies stood up in August. The Department of War’s Economic Defense Unit will identify small businesses that need financing and refer loan candidates into the SBA’s 7(a) and 504 lender networks. The unit may also originate Department-funded redeemable preferred-equity investments in firms that create, expand or restore industrial capacity judged critical to national security, with the SBA running due diligence, underwriting, deal structuring and portfolio management while the unit keeps final decision authority. Secretary of War Pete Hegseth said the commission “will now be able to surge strategic capital,” and SBA Administrator Kelly Loeffler put small manufacturers at 70% of the defense industrial base. Worth reading twice before filing it under good news: capital is becoming available to expand capacity, and the companies that take it will be bidding for the same controlled technical data you handle. Money for machines is not money for your affirmation.
  • Oct 6, 2026
    The GSA schedule’s clause list is about to get shorter — and the AI clause is riding in with it.
    Wave 4 of GSA’s acquisition-regulation overhaul, folded into the same November refresh, deletes 41 existing provisions and clauses, removes eight legacy schedule clauses and consolidates five replacements at 552.238-190 through -194. One cybersecurity change worth diarizing separately: the incident response subgroup under the highly adaptive cybersecurity services category is being phased out at the end of FY2027, with holders pointed to the incident handling and event management subgroup instead. If you maintain a clause matrix spanning a DFARS-covered contract and a civilian schedule, that matrix changes twice this quarter.

Not sure where your organization stands?

CUI Vault gives you a live, always-audit-ready picture of your CUI environment and NIST 800-171 posture — so headlines like today’s are a footnote, not a fire drill. Reach out to your NtelSec team to talk through your current gaps.

This briefing is a general-information summary of publicly reported CMMC and CUI developments. It is not legal advice; consult your compliance counsel before making contractual or certification decisions.
Talk to us about CUI Vault All issues