Protecting Enterprise Data in AI Workflows
A company deciding whether to send source code, legal documents, or research through a generative AI service needs more than a training-policy summary. It needs to know who can access inputs, what data the service collects, and how those commitments are enforced. VentureBeat’s coverage of OpenAI’s DevDay 2026 announcements brings those questions into focus.[1]
Enterprise AI Data Concerns and Withdrawals
VentureBeat, summarizing reporting by The Information, names Palantir, Nvidia, and Booz Allen Hamilton among companies that have reportedly restricted or reconsidered use of advanced AI models. The coverage does not identify one shared reason for their decisions, and it does not report that any of the companies experienced a breach.[1]
VentureBeat reports that OpenAI and Anthropic say they do not train on information covered by enterprise agreements by default, unless customers opt in. That commitment does not necessarily rule out other collection: the report notes that technical usage metadata may still be gathered.[1]
The report also relays figures disclosed by OpenAI: work files in ChatGPT were reportedly accessed more than a million times a day, and about 0.02% of conversations were flagged for further safety review. OpenAI said that, in a very small number of flagged cases, users saw outputs that appeared to draw on other users’ enterprise work files.[1] These are OpenAI’s reported figures and account of the cases. VentureBeat does not establish how they occurred or confirm that an outsider obtained company data.
The coverage does not establish that data-handling concerns caused any particular company to change its AI use. It also discusses a separate Navier–Stokes mathematics controversy, without connecting it to enterprise data handling.[1]
OpenAI’s Answer: Private Intelligence
At DevDay 2026, OpenAI announced “Private Intelligence,” a two-part initiative, according to VentureBeat. The first part, a privacy-focused safety-review system, was available at the time of the announcement. The second, a confidential-computing architecture, was planned for later.[1]
The available component is Zero Data Retention with Private Safety Processing (ZDR with PSP). OpenAI describes it as a way to run safety checks without retaining customer data. The company says customer information stays in the customer’s environment and is protected by customer-managed authorization keys. In the design VentureBeat describes, a safety-review application reads from a queue and sends inputs through an automated safety layer; customer context remains in that environment before the model call.[1]
OpenAI says the system uses hardware attestation and SOC 2 auditing, excludes human review of customer content, and returns classification results—not raw inputs and outputs—to the company. The announced flow also includes post-generation review: a block-or-allow decision returns to the queue, with alerts for violations.[1] These details describe OpenAI’s stated design; they are not independent evidence of implementation or effectiveness.
The second component, Private Inference, was previewed, not shipped. OpenAI said it would use confidential computing to extend confidentiality to model execution and expected to make it available later that fall.[1]
VentureBeat’s account leaves practical details unresolved: which models and API endpoints would support Private Inference, whether it would cover ChatGPT Enterprise as well as API inference, what trusted-execution hardware it would use, what customers could verify independently, and what it would cost.[1] Those specifics will determine how well the offering fits a buyer’s needs.
What Anthropic Offers, and Where Rivals Fit
Anthropic’s API “zero retention” option was in beta for business-to-business customers, VentureBeat reports. Anthropic’s stated terms say prompts and outputs are processed transiently to generate a response, then neither stored nor used for model training. The company also says human reviewers do not examine customer inputs and outputs under this arrangement.[1] These are Anthropic’s claims, not independently verified findings.
VentureBeat also notes Microsoft’s emphasis on data visibility in its marketing of private AI infrastructure, alongside confidential-computing options from other hyperscalers.[1] The report does not suggest that these offerings are interchangeable; their protections depend on the service and its configuration.
Contract, Architecture, and the Metadata Gap
A contract sets obligations; technical measures may limit access in a particular configuration. For enterprise buyers, the important distinction is what each commitment covers and what evidence is available to verify it.
Metadata is one place where a promise not to train on customer content may leave questions unanswered. VentureBeat’s summary of The Information says telecommunications operator C Spire has agreements with OpenAI and Anthropic that prohibit training on its data while permitting collection of technical usage data.[1]
“Metadata can inadvertently reveal sensitive information about our operations,” Michael McKenzie, C Spire’s vice president of data and analytics platform engineering, told The Information, as quoted by VentureBeat. “We can’t fully prevent our data from helping improve [the providers’] models.”[1]
McKenzie’s remarks convey his concern; they do not establish that metadata from every service reveals operational details or is used to train models. The distinction makes the provider’s definition and use of metadata worth examining alongside its content-retention and training terms.[1]
Questions for an AI Provider
The reported pullbacks and the providers’ stated controls leave buyers with specific points to clarify before signing:
- Does zero retention cover inputs, outputs, and intermediate traces, or is any content retained for safety review?
- Are flagged items reviewed by automated systems, human reviewers, or both?
- Can the customer control the keys and keep content in its own environment? What hardware-attestation evidence can it inspect?
- What counts as permitted metadata, how is it used, and can collection be limited?
- Do the protections cover both API inference and the enterprise chat workspace employees use?
What Has Actually Been Breached, and What Has Not
VentureBeat’s report does not describe a publicly confirmed case of an enterprise customer losing data from a provider-side system.[1] That is a limit of what this reporting establishes, not proof that no incident has occurred. Nor do the reported decisions by Palantir, Nvidia, and Booz Allen Hamilton establish that those companies experienced a breach or shared the same reason for changing their AI use.[1]
For security leaders, the useful comparison is between each provider’s contractual commitments, stated technical controls, and permitted data collection. The terms and implementation matter; metadata practices should be checked service by service rather than assumed to be uniform.
Sources
- [1] Carl Franzen, “OpenAI’s new ‘Private Intelligence’ targets a growing enterprise concern: who can see your AI data?” VentureBeat, September 29, 2026: https://venturebeat.com/security/openais-new-private-intelligence-targets-a-growing-enterprise-concern-who-can-see-your-ai-data. VentureBeat also summarizes The Information’s reporting concerning Palantir, Nvidia, Booz Allen Hamilton, C Spire, Michael McKenzie, enterprise training policies, and metadata collection. Claims drawn from those details are attributed above to VentureBeat’s summary.