OpenAI launches Private Intelligence: encrypted safety-review records stay in customer-owned cloud storage with no human access, and confidential-computing Private Inference enters API preview

Why it matters

【Key points】

On 2026-10-30 (15:12 UTC) OpenAI launched a dedicated page for OpenAI Private Intelligence, an enterprise privacy package for using its most capable models, built with "some of our largest customers". (1) ZDR with Private Safety Processing: classifier-referred or sampled records are double-encrypted and written to customer-owned storage (AWS S3, Azure Blob, GCS or OCI) with a 30-day TTL. They are decrypted only in a hardware-attested runtime with no human access, and optional customer-managed EKM keys can block decryption. (2) Private Inference: confidential computing with "strict, verifiable controls" during processing, labeled an API preview for eligible workflows (register interest). It also covers data ownership, agent permissions and approvals, SOC 2 Type 2, ISO, HIPAA BAA and FedRAMP, across ChatGPT, Codex and the API. The name first appeared at DevDay on Sep 29. No pricing, model list, hardware, GA date or named customers have been published.

【Verifiable】

https://openai.com/business/private-intelligence/ ; https://developers.openai.com/api/docs/guides/private-safety-processing ; https://x.com/OpenAI/status/2116186509384815084

【Our read】

PSP solves the tension between cross-interaction safety review and zero retention by keeping the data in the customer's own bucket, which matters for regulated and sovereignty-minded buyers. Peers such as Azure OpenAI data zones and CMEK, Claude Enterprise ZDR/CMEK, and Apple Private Cloud Compute cover parts of this. OpenAI's twist is putting safety review itself inside confidential computing. Caveats: "zero retention" now means selected records stay in your own bucket for at least 30 days, and PSP applies to a project's entire API traffic. Until hardware, attestation details and an independent audit appear, Private Inference is a roadmap, not a guarantee.