How to Keep Sensitive Facility and Security Drawings Out of AI Tools

Which drawings need special handling, from SSI drawings to client security plans?

Any drawing a law, a government contract or a client agreement restricts. For most AEC firms that means three groups, and a project can carry more than one.

• Sensitive Security Information (SSI). On airport, port and other transportation work, TSA's rule at 49 CFR part 1520 covers items such as security programs and contingency plans, vulnerability assessments, performance specifications for screening and detection equipment, and lists of critical infrastructure assets. Contractors and their employees are covered persons under section 1520.7.
• Federal building information marked as CUI. GSA's order PBS 3490.3 CHGE 1 treats qualifying information about GSA-controlled buildings as Controlled Unclassified Information in the Physical Security category and directs all holders to protect it under 32 CFR part 2002. Not every federal building drawing is CUI; the marked ones are.
• Defense and private security work. CUI on defense projects falls under DFARS 252.204-7012. Private clients such as data centers, banks and labs may restrict security layouts, camera and access control plans, and MEP details through contract and NDA terms.

Why is uploading these drawings to an AI tool a problem?

Because an AI service outside your controls is a new recipient, and these rules limit who may receive the information. The upload can be an unauthorized disclosure on its own, whether or not anything bad happens afterward.

Under 49 CFR 1520.9, a covered person must take reasonable steps to safeguard SSI and may disclose it only to covered persons with a need to know. If SSI reaches an unauthorized person, the covered person must promptly inform TSA or the applicable DOT or DHS component, and section 1520.17 makes violations grounds for civil penalties and corrective action.

For CUI, 32 CFR 2002.14(h)(2) points agencies to NIST SP 800-171 for non-federal systems, and NIST SP 800-171 Rev. 2 requirement 3.1.22 calls for controlling CUI posted or processed on publicly accessible systems. A public chatbot or free AI rendering site is a publicly accessible system. On defense work, DFARS 7012 also requires a cyber incident report to DoD within 72 hours of discovery.

How do drawings usually end up exposed?

Ordinary sharing settings and convenience tools are enough to expose them. A GSA Office of Inspector General alert memorandum dated April 18, 2025 found sensitive building information in GSA's Google Drive that all GSA users, including contractors, could view. The exposed files included a U.S. courthouse feasibility study with blueprints showing mechanical rooms, grand jury rooms, U.S. Marshals Service offices and the server room, and nine documents marked CUI, among them a site development plan with fire protection and MEP details and electrical drawings of the main switchgear.

AI tools add new paths of the same kind. Common ones on design teams include:

• Pasting specifications or uploading sheets to a public chatbot to summarize, check or rewrite them.
• Running plans or renderings through free AI upscaling and visualization websites.
• AI features vendors add inside PDF markup, document management and design software.
• AI notetakers recording design reviews while security drawings are on screen.
• Browser extensions that read every page the user opens.

How do you keep restricted drawings out of AI tools?

Decide at project intake, then let the systems enforce the decision. Controls a small or midsize firm can run:

1. Classify at intake. Have the project manager check the contract, the RFP and any markings for SSI, CUI, security or confidentiality clauses, and flag the project as restricted in your project system.
2. Use a restricted project space. Keep restricted files in a separate library or folder with named members only, no anyone-with-the-link sharing, and no external sharing without approval.
3. Label the files. Apply a sensitivity label with encryption to restricted projects. Microsoft states that Microsoft 365 Copilot only surfaces data a user can already view and honors the usage rights set by Purview encryption, so permissions and labels decide what an approved AI assistant can reach.
4. Block uploads to unapproved AI. Data loss prevention can stop labeled files and marked text from reaching unmanaged AI sites. Microsoft documents a Purview policy for unmanaged AI apps in Edge for Business that blocks both pasted text and file uploads on managed devices.
5. Filter sites and extensions. CIS Controls v8.1 Safeguards 9.3 and 9.4 call for URL filtering of unapproved sites and restricting unauthorized browser extensions, which covers free AI tools and AI add-ons.
6. Turn off AI features on restricted work. Where design, markup or meeting tools allow it, disable AI features for restricted projects, and keep notetakers out of security design reviews.
7. Flow it down. Put the same AI rules in subconsultant agreements and confirm how consultants will store your restricted files.

On defense work, these controls overlap with NIST SP 800-171. Our article on which NIST 800-171 gaps to close first covers the order of work.

What should you do if a restricted drawing already went into an AI tool?

Treat it as an incident and move quickly, because some of these rules set short reporting clocks. Record who uploaded what, to which tool, when and under which account, and preserve that record before anyone deletes the chat.

Then work out which rules apply. SSI requires prompt notice to TSA or the relevant DOT or DHS component. CUI on a defense contract can trigger the DFARS 7012 72-hour report. Client contracts may carry their own notice terms. Deleting the conversation in the AI tool is worth doing under the vendor's process, but it does not undo the disclosure, so make reporting decisions with counsel and the client's contract in hand.

What should you do this quarter?

Start with the projects you already have:

1. List active and pursued projects with SSI, CUI, federal building or security-sensitive scope.
2. Move those files into restricted spaces and remove broad or link-based sharing.
3. Apply sensitivity labels and turn on DLP for uploads to unapproved AI sites.
4. Review AI features in your design, markup and meeting software and switch them off for restricted work.
5. Add a restricted-project briefing to kickoff meetings and update subconsultant terms.
6. Write a one-page incident procedure covering who to call and which notices may apply.

How does s90 help?

s90 supports architecture, engineering and construction firms with the controls described here: restricted project spaces, sensitivity labels, DLP and AI guardrails that keep security drawings in approved systems, and zero trust access for staff and consultants. If you have security-sensitive projects coming up, we can walk your team through what a restricted project setup looks like.