OpenAI introduced Private Intelligence at its DevDay event Tuesday, a program meant to limit what the company itself can see of enterprise data that passes through its models, VentureBeat reported. The first piece, available now to eligible API customers, stores safety-flagged content encrypted in the customer's own cloud account instead of on OpenAI's systems.

What Happened

Private Intelligence has two parts. The first is Zero Data Retention with Private Safety Processing. Under it, customer content selected by a safety classifier or an approved sampling policy is encrypted and written to storage the customer controls, in Amazon S3, Azure Blob Storage or Google Cloud Storage. OpenAI keeps operational metadata and storage references but no plaintext copy.

When a record needs safety review, it is decrypted only inside what OpenAI calls a hardware-attested safety runtime, which is designed so that OpenAI employees cannot view the underlying content. Predefined safety signals and metadata come out in plaintext; detailed results remain encrypted. The records carry a 30-day time-to-live, and the authorization model supports customer-managed Enterprise Key Management. According to VentureBeat, protected material cannot be used to train models or be shared with OpenAI partners.

RelatedNew York City Comptroller Urges Anthropic to Delay IPO Over Safety and Governance Concerns

The second part, Private Inference, is due in preview in fall 2026. OpenAI said it will pair confidential computing with verifiable controls while data is being processed, but did not name the models, hardware, pricing or verification methods involved. No pricing was given for the zero-retention option either.

Why It Matters

Abuse monitoring has been a sticking point for regulated buyers. Providers review flagged traffic for misuse, and any copy held for that purpose sits outside the customer's control, which is what compliance and security teams tend to object to. OpenAI's approach keeps the safety review but moves the stored copy into infrastructure the customer owns and can audit, with the customer's keys controlling access.

The label needs careful reading. VentureBeat pointed out that "zero data retention" here describes who controls and can access the data, not the absence of any copy: encrypted records do exist, in the customer's bucket. Customers also take on work, since they must keep the storage environment and key authorization available for the full 30-day window.

The Trust Gap OpenAI Is Targeting

The launch responds to hesitancy among large customers. VentureBeat reported that companies including Palantir, Nvidia and Booz Allen Hamilton had begun restricting model use or pressing for stronger contractual guarantees, amid concern that OpenAI or Anthropic could learn from proprietary information. Telecom provider C Spire, the outlet noted, has agreements with both labs that bar training on its data while still allowing collection of technical usage data.

Rivals are pitching the same concern. Microsoft has used enterprise data protection as a selling point for its private AI infrastructure, VentureBeat said, and hardware-backed confidential computing has become a common answer from cloud providers to the question of who can see data in use.

What to Watch

The practical test is Private Inference, which would extend protection from stored safety records to the processing itself. Buyers will want to know which models it covers, what hardware attestation customers can verify on their own, and what it costs. Pricing for both components, and whether eligibility widens beyond the current set of API customers, remain open.