Does Your AI Vendor’s Contract Protect Your Trade Secrets? Read Past the DPA

Terms as of Oct 20, 2026; vendors update terms often. Enterprise Unified Solutions (EUS) provides technical and vendor-risk assessments of AI tools and their contracts. EUS is not a law firm, and this article is not legal advice. Review contract, trade-secret, export-control, tax and ethics questions with your own attorney.

This post is for IT and compliance leads at small to mid-market manufacturing, distribution, professional services, or defense-supplier organizations across the USA including Indiana, western Ohio and Florida. Specifically, BLS Size Class 4 (20 to 49 employees) to Size Class 8 (500 to 999 employees).

The short answer: partly, and only on the tier you signed

The major AI vendors’ enterprise terms now say they won’t train on your content and that you own your inputs and outputs. Most treat it as confidential. Consumer plans (free or individual paid accounts) generally don’t.

The catch is the document buyers check first. A DPA (data processing agreement, or addendum) is written for privacy law, so it mostly governs personal data. Your source code, CAD files, pricing and client workpapers are protected, if at all, by the main agreement’s confidentiality and data-use clauses. Gaps remain around retention, abuse monitoring, subprocessors, liability caps and staff on consumer accounts. The contract is necessary, not sufficient.

What the enterprise terms say

Vendor (controlling documents)Trains on your content?What breach notice coversNotice timingLiability cap
OpenAI (Services Agreement and DPA, eff. Jan 1, 2026)No, unless you agreePersonal dataWithout undue delay12 months’ fees
Microsoft (Product Terms; DPA, May 22, 2026)Not for foundation modelsAll Customer DataWithin 72 hoursIn your licensing agreement
Google Workspace (terms, Oct 12, 2026; Cloud DPA, Jun 8, 2026)Not without permissionCustomer DataPromptly, without undue delay12 months’ fees
Anthropic (Commercial Terms, Jun 17, 2025; DPA, Feb 24, 2025)NoPersonal dataWithin 48 hours12 months’ fees
Cursor (Terms, Sep 3, 2026; DPA, Nov 5, 2025)No, unless you agreePersonal data, where privacy law requiresWithin 48 hoursOnline terms: greater of 6 months’ fees or $100 (an Enterprise MSA governs if signed)

Definitions decide coverage. OpenAI’s DPA defines Customer Data as “Personal Data processed by OpenAI on behalf of Customer.” Anthropic’s and Cursor’s DPAs tie breach notice to personal data. Microsoft’s DPA is broader: Customer Data means “all data, including all text, sound, video, or image files, and software” you provide. Google’s Cloud DPA also covers Customer Data.

On confidentiality, OpenAI, Google and Anthropic each define customer content as your Confidential Information. Anthropic’s commercial terms put training plainly: “Anthropic may not train models on Customer Content from Services.” In our reading, Cursor’s public Terms of Service contain no vendor confidentiality promise over your content, so an Enterprise MSA (not public) needs to supply one. Microsoft’s cap and confidentiality terms sit in your licensing agreement; read it directly.

Vendors also rely on other AI companies. Microsoft’s Copilot documentation names Anthropic and OpenAI as subprocessors (outside firms that process your data for the vendor). Cursor’s maker, Anysphere, is now a SpaceX subsidiary, according to Cursor and an SEC filing, and its Data Use page lists SpaceXAI among its model providers with OpenAI, Anthropic and Meta. You don’t sign those downstream contracts and usually can’t see them.

Where the DPA stops: our analysis

EUS’s readings, not legal conclusions:

  • Abuse monitoring (vendor retention and possible human review of content to detect misuse). OpenAI’s API documentation keeps abuse logs up to 30 days by default; zero data retention (ZDR), where nothing is stored after the response, needs OpenAI’s approval. Microsoft’s Azure OpenAI documentation allows sampled human review unless modified monitoring is approved, while Microsoft says Copilot has opted out. ZDR is granted, not switched on.
  • Liability caps (the ceiling on what the vendor owes you). None we read carves out confidentiality breaches, so a secret worth millions may sit behind months of fees.
  • Connectors and web search. Microsoft’s Copilot documentation says web queries go to Bing, outside the DPA and BAA. OpenAI’s residency page says data handled through integrations may be stored outside your chosen region.
  • Change by posting. Cursor’s online terms change when posted, Anthropic’s 30 days after posting, and OpenAI’s on notice (30 days if material), though signed OpenAI Order Forms are protected. Google Workspace defers material changes to renewal.
  • Export-controlled data. OpenAI’s agreement states that “Customer’s Input may not include material or information that requires a government license for release or export.” The ITAR and EAR exemption for encrypted transfers requires that the means of decryption not go to any third party (22 CFR 120.54; 15 CFR 734.18). Our inference, for export counsel to confirm: a hosted model must decrypt prompts, so the exemption likely doesn’t cover inference.

Reasonable measures: the contract is only half

Under the federal Defend Trade Secrets Act, 18 U.S.C. §1839(3), information is a trade secret only if the owner has taken “reasonable measures” to keep it secret. In our reading, that test reaches what your staff actually do, including which AI terms they click through.

The risk is behavioral. In April 2023, The Register, relaying Economist Korea, said Samsung had “reportedly leaked its own secrets into the AI service at least three times,” including semiconductor source code and meeting records. Samsung later banned generative AI on company devices. No contract helps when someone pastes into a personal account.

Professional duties point the same way

ABA Formal Opinion 512 asks lawyers to read a tool’s terms and notes they “may need to consult with IT professionals or cyber security experts to fully understand these terms and policies.” The California State Bar’s guidance says diligence may include “consulting with qualified IT professionals or cybersecurity experts.” Marketing assurances alone aren’t enough.

CPA firms have IRC §7216, making knowing or reckless disclosure of tax return information a crime. Treasury regulations add consent requirements when the recipient preparer is outside the US. Whether an AI vendor counts is unsettled; ask tax counsel.

A BAA (business associate agreement, the HIPAA contract a vendor signs before handling patient data) covers only named products. Microsoft excludes web search, Google excludes Gemini Notebook, and Cursor’s BAA does not automatically cover third-party integrations.

Defense suppliers: the CMMC pause didn’t pause 7012

On July 13, 2026, DoD suspended CMMC Phase II, which covered C3PAO (third-party assessor) assessments (Peckar & Abramson; Cyber AB). DFARS 252.204-7012 and NIST SP 800-171 Rev 2 still apply. DFARS 252.204-7012 requires cloud providers handling covered defense information to meet FedRAMP Moderate-equivalent security and to support incident reporting “within 72 hours of discovery.” In our reading, none of the five vendors’ standard AI terms passes those duties through to you, so AI touching covered defense information needs a suitable offering and a matching contract.

Ten questions to ask your vendor

Our analysis; counsel should finalize any redline wording.

  1. Which document governs our tier, and can it change without our signature?
  2. Is all our content, including outputs, embeddings and logs, our Confidential Information?
  3. Does anyone, including subprocessors and model providers, train or evaluate on our content?
  4. What is retained, how long and where, including abuse logs, telemetry and backups?
  5. Can employees or contractors see our prompts? When, and from which countries?
  6. Who are all subprocessors and model providers, what notice do we get, and can we object?
  7. Where do storage, inference and support staff sit?
  8. Does breach notice cover all our content, not only personal data, and within how many hours?
  9. Does the liability cap carve out confidentiality and data breaches?
  10. At exit, when is everything deleted, and will you sign a deletion certificate?

Enterprise no-training covenants are generally adequate; confirm them in your signed copy. Questions 1, 2, 8 and 9 are the usual redlines. Questions 4 through 7 often come down to settings and approvals.

Controls that don’t live in the contract

Some risk only technical controls reduce. Block consumer AI domains and personal-account logins on managed devices. Require SSO and MFA, with automated deprovisioning when employment ends. Lock admin settings: training opt-out, ZDR or privacy modes, retention, public sharing off, connectors allow-listed. Label sensitive data and fix oversharing before AI indexes it. Add DLP (data loss prevention) for pastes and uploads. Microsoft Purview is one option; if your files live in Egnyte, Egnyte says its AI Safeguards can keep sensitive content out of AI responses, and it integrates with DLP tools including Purview. Publish an approved-tools list by data class, backed by an acceptable-use policy and training. Re-review terms yearly. Add AI exposure to your incident response plan. For the most sensitive data, consider private hosting so no outside model provider holds the keys. Even then, residual risk may not be eliminated.

Where EUS fits

EUS offers an AI DPA and contract technical assessment for protection of company secrets: a technical review of your AI vendor’s terms and data flows that works alongside your counsel, not a legal review. EUS also partners with Egnyte and can implement and manage its governance, DLP and AI Safeguards controls. EUS is a CMMC Registered Practitioner Organization (RPO).

Sources