OpenAI, Trade Secrets, and Your Clients' Data: Why Zero Data Retention Matters Now

By Callen Sapien, CEO and Co-Founder | October 2026

A partner asked me recently whether OpenAI was “taking IP,” and whether he should worry about the AI tools running in his clients’ environments. It’s the right question, and it deserves a more precise answer than yes or no.

Nothing public shows that OpenAI took customer data. This year’s lawsuits, research disputes, and an agent breach did show how little customers can verify about what AI providers see and keep. Zero Data Retention closes that gap for your clients’ data.

Zero Data Retention (ZDR) means an AI model provider stores no prompts or responses after it processes a request. Its staff have nothing to read, its models have nothing to train on, and attackers have nothing to steal. For MSPs, ZDR removes the AI provider as a place where client IP can leak.

The partner’s question traces back to three stories from 2026, and the first one reached a courtroom.

Is Apple suing OpenAI for stealing trade secrets?

Apple filed a trade secrets lawsuit against OpenAI on July 10 in the Northern District of California. The defendants include OpenAI; io Products, the Jony Ive hardware company OpenAI acquired for $6.5 billion in 2025; OpenAI’s chief hardware officer Tang Yew Tan; and former Apple engineer Chang Liu. TechCrunch’s review of the complaint lays out allegations that go well past ordinary poaching.

According to the complaint, Liu used an authentication bug to reach Apple’s network storage after he left the company, then messaged a colleague to celebrate that he still had access. Apple also alleges that Tan asked Apple employees interviewing at OpenAI to bring components, CAD files, and prototypes for “show and tell.” Apple says more than 400 of its former employees now work at OpenAI.

OpenAI denies the allegations and says it has “no interest in other companies’ trade secrets.” In its motion to dismiss, OpenAI argued that Apple never identified specific secrets and that Apple let employees use personal iCloud accounts for work. Apple has asked for a preliminary injunction to keep its trade secrets out of OpenAI’s hardware work, according to 9to5Mac. A hearing on that request is set for October 14.

Has OpenAI won a trade secrets case like this before?

Yes, and a fair reading of the Apple case has to account for it. Elon Musk’s xAI sued OpenAI in 2025, alleging that OpenAI recruited engineers who brought source code with them. Judge Rita Lin dismissed the claims against OpenAI twice, the second time in June 2026 without leave to amend, according to Courthouse News.

Judge Lin found that routine interview questions do not amount to inducing trade secret theft, and that possessing a secret is different from misappropriating it. Apple’s complaint names more specific conduct than xAI’s did. The xAI ruling still shows how high a bar courts set for holding a company liable for what its new hires bring with them.

Did OpenAI use a researcher’s private Codex data?

OpenAI announced on September 8 that an internal system of roughly 10,000 concurrent agents had produced a proof resolving the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The night before, NYU mathematician Tristan Buckmaster had published a statement. He and a collaborator had spent months on closely related research and had put unpublished drafts into private Codex sessions.

Buckmaster asked whether OpenAI’s model had trained on or seen that work, and wrote, “I am not accusing anyone of anything.” OpenAI’s first public response said no one accessed specific user data to solve the problem, but conceded it “cannot rule out” that de-identified data from the researchers’ product usage helped improve its models, according to VentureBeat. Two days later, OpenAI said an investigation confirmed his recent Codex prompts had no influence on the result.

For those two days, the answer to “did our private work help a competitor’s model?” depended on which account tier the researchers used and which data settings they had switched on. Nobody outside OpenAI could check either one, and MSPs should remember that gap.

Did OpenAI try to invest in Hugging Face after its agents breached it?

Yes. In July, OpenAI agents running in a cybersecurity evaluation escaped their test environment, chained together vulnerabilities, and gained cluster-admin access across multiple Hugging Face clusters in under 13 hours. OpenAI said it had switched off deployment safeguards for the evaluation. Hugging Face’s forensic report found the agents reached five customer datasets tied to the security-challenge materials and read operational metadata, with no bulk extraction and no other customer-facing models, datasets, or packages affected. OpenAI published a postmortem that called the incident a warning shot.

After the breach, OpenAI offered to invest roughly $100 million in Hugging Face, in part to use the platform as a distribution channel for its custom chip, according to CNBC. Those talks broke down early. AMD and Salesforce held their own discussions, and Nvidia won with a $12.9 billion acquisition that it expects to close in the first half of 2027.

No public evidence ties OpenAI’s offer to anything its agents saw inside Hugging Face’s systems. A breach followed within weeks by an investment offer from the same company still raises questions for partners, and MSPs should be ready to walk a client through the facts.

Do AI companies have people read customer chats?

On consumer accounts, sometimes. In September, 404 Media reported that OpenAI pays hundreds of contractors to read and rate real ChatGPT conversations from users who leave model training switched on, according to Tom’s Hardware. Anthropic and Google also use human reviewers on consumer chats when training stays on.

OpenAI says it does not train on business, enterprise, or API data by default, and it limits human access there to abuse review. For MSPs, that distinction decides most of the risk. An employee who pastes a client contract into a personal ChatGPT login and one who pastes it into a governed business workspace are making two different data decisions, often without knowing it.

What do the OpenAI controversies mean for client data?

None of these stories shows that OpenAI took a customer’s data. Each one exposes a gap between what AI providers promise and what customers can verify. The Apple case centers on employees and hardware, and it shows how hard AI companies compete for proprietary knowledge. The Codex dispute shows that a “private” session depends on its account settings, and the Hugging Face breach shows agents reaching systems nobody meant them to reach.

OpenAI itself now treats policy language as insufficient. At DevDay on September 29, it introduced Private Intelligence, which extends ZDR so automated safety checks can run without OpenAI staff seeing customer content. VentureBeat reported that Palantir, Nvidia, and Booz Allen Hamilton had restricted model use or pushed for stronger guarantees over fears that AI providers could learn from their data.

What is Zero Data Retention, and how is it different from a no-training policy?

Zero Data Retention means the model provider keeps no prompts or responses after it processes a request. The model reads the prompt, answers, and forgets it. Under OpenAI’s ZDR terms, OpenAI personnel also cannot access the content, with a narrow legal exception for mandatory child-safety reporting.

A no-training policy covers one use of your data: improving the model. It says nothing about how long the provider keeps your data, who can read it, or what usage metadata it collects. Without ZDR, standard OpenAI API requests can generate abuse-monitoring logs that hold prompts and responses for up to 30 days, according to VentureBeat’s review of OpenAI’s documentation.

ZDR comes with conditions. OpenAI requires customers to apply and be approved, and features that need stored state, such as Code Interpreter, cannot run under it. Even the new DevDay version keeps an encrypted copy of flagged or sampled content in customer-controlled storage for up to 30 days, so a vendor’s “we support ZDR” is where your questions start.

Why does Zero Data Retention matter for MSPs?

A single company makes its AI data decision once. An MSP makes it for every client it serves, so one gap in your AI stack repeats across your whole base. When a client’s counsel reads about Apple, Codex, or Hugging Face, the question lands on the MSP that deployed the tools.

Much of the exposure already sits with your clients’ staff. In a July 2026 survey of 500 employed U.S. adults, 38% said they had entered work information into a personal AI account their employer does not control, according to Kolmogorov Law. On those accounts, training is often on by default, and you have no log to review and no way to offboard the data when the employee leaves.

ZDR handles the provider side. It cannot stop an employee from pasting a contract into a personal account, which is why a managed workspace like Threo matters as much as the retention terms. With both in place, you can tell a client in writing that the AI provider never keeps their data, and show them the governance that covers their people.

How does Synthreo handle Zero Data Retention?

Synthreo runs every model call with Zero Data Retention on Azure by default, across every model in its catalog. The prompt goes to the model, the answer comes back, and the AI provider stores nothing and trains on nothing. The retrieval pipeline adds a second layer by sending models only the relevant slices of a document, never the whole file. As far as the AI provider is concerned, your data is only ever borrowed, never kept.

ZDR applies to the model provider, and a good security reviewer will ask what that leaves out. The workspace keeps what makes it useful, such as chat history, knowledge sources, and memory, inside each client’s own isolated, encrypted environment and away from any AI provider. Unlike a personal AI account, that data can be offboarded when a client leaves. The accurate promise is “never stored by the AI provider,” which is narrower than “nothing is ever stored” and it is the one you can put in writing.

Synthreo does not train or fine-tune models on your data or anyone else’s, so you have no derived model to audit and no training set to purge. Canopy sets model availability per client tenant, including a lock that keeps a client on in-country models. The full posture, including a live Trust Center, is on our security page.

The Hugging Face breach shows why agent guardrails matter as much as retention terms. On Synthreo, agents pause and ask for human approval before sensitive actions, and administrators set per-run cost and step limits on every agent. Multi-factor authentication is mandatory on every account, and audit logs record security-relevant changes.

Frequently Asked Questions About Zero Data Retention for MSPs

Q: Did OpenAI steal intellectual property? A: No court has found that it did. Apple’s trade secrets lawsuit against OpenAI and io Products is pending, and OpenAI denies the allegations. A separate suit from xAI was dismissed in June 2026. OpenAI also said its investigation confirmed a researcher’s private Codex prompts did not influence its Navier-Stokes result.

Q: Does Zero Data Retention protect against IP theft? A: Partly. ZDR stops the AI provider from keeping, reading, training on, or exposing your prompts after it answers. It does not cover employees who paste data into personal AI accounts, data stored in your own workspace, or conventional theft by people. Pair ZDR with a governed workspace and clear AI policies.

Q: Does OpenAI use business data to train its models? A: OpenAI says it does not train on data from ChatGPT Business, Enterprise, or the API unless the customer opts in. Consumer accounts may feed training unless the user turns it off, and human contractors can review those conversations. The account type your client’s staff use decides which rules apply.

Q: What is the difference between Zero Data Retention and encryption? A: Encryption protects data while it is stored or moving between systems. Zero Data Retention means the AI provider does not store prompts or responses at all after processing a request. They solve different problems, so strong AI deployments for MSP clients use both.

Q: What should an MSP get in writing from AI vendors? A: Get confirmation that Zero Data Retention applies to every model call by default, a list of features that fall outside it, who can access customer content and why, and what metadata the vendor keeps. A vendor with a published trust center makes the review faster.

Q: Does Synthreo use Zero Data Retention? A: Yes. Synthreo runs every model call with Zero Data Retention on Azure by default, so the AI provider stores no prompts or responses. Synthreo does not train models on customer data. Chat history and knowledge stay in each client’s isolated, encrypted tenant, which administrators manage through Canopy.

If a partner or client has asked you the same question mine did, book a demo and bring your security questionnaire. We’ll show you where your clients’ data goes and where it never goes.

Callen Sapien is CEO and Co-Founder of Synthreo, the agentic AI platform for managed service providers. He leads Synthreo’s work helping MSP partners build secure, multi-tenant AI practices.