OpenAI’s privacy push puts safety checks inside a customer-controlled boundary

OpenAI has announced two privacy measures for AI workloads: an enhanced safety-review system available now and a confidential-computing service still in preview. The practical change is that automated safety checks can run without OpenAI staff seeing the customer content being checked.

OpenAI Watch analysis

What happened

VentureBeat reports that OpenAI’s Private Intelligence initiative combines Zero Data Retention with Private Safety Processing, or ZDR with PSP, and a forthcoming service called Private Inference. OpenAI says ZDR with PSP is available now to eligible API customers. Private Inference is due this autumn, according to the report.

With ZDR with PSP, selected prompts and responses are encrypted and stored in a customer-controlled AWS, Azure or Google Cloud environment for up to 30 days. OpenAI says an attested safety runtime can retrieve and review them automatically, while OpenAI staff cannot inspect the underlying content. The system is designed to return limited safety signals and approved operational metadata, rather than readable customer material.

Why it matters

Zero data retention does not mean that no copy exists anywhere: under this arrangement, encrypted records remain in customer-controlled storage temporarily so automated safety review can happen. Customers must maintain the storage, permissions and any key authorisation they use. That distinction will matter to teams assessing their own retention and security obligations.

OpenAI says material handled through PSP cannot be used to train models or shared with other OpenAI groups or partners. Private Inference is less defined so far: the report says OpenAI has not yet detailed which models or endpoints it will support, or what customers will be able to verify independently.

Our read

This is a meaningful attempt to make a privacy promise more technical than “trust us”. It gives eligible API customers a route to automated safety checks while keeping protected content out of OpenAI staff’s view, according to the company’s description in VentureBeat’s account. But the customer still has storage and access controls to manage, and the preview of Private Inference is not yet a working specification. Security teams should distinguish what is available now from what is promised for later; architecture diagrams are not a substitute for deployment details.

What to watch

• Which API endpoints and models support ZDR with PSP, and what customers need to configure.
• What Private Inference will cover and how customers can verify its controls.
• Whether OpenAI publishes enough technical detail for enterprise security teams to assess the design.

Discussion spark: For sensitive AI workloads, is customer-controlled encrypted storage with automated safety review a useful compromise, or does any temporary copy remain a deal-breaker?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.