OpenAI has begun previewing a feature called Private Safety Processing, designed for customers who run its models under a zero-data-retention policy. The move signals that the company is trying to close a long-standing gap: how to apply safety filters without keeping any record of the prompts that triggered them.
What the preview covers
For organizations that sign up for zero data retention, OpenAI typically promises that prompts and responses are not stored. But safety checks have historically required some form of logging or inspection, which created a conflict. The new preview appears to let those checks run in a way that doesn't leave a permanent trace.
The company hasn't shared technical details about how the processing is isolated or what guardrails apply. It only says the feature is being previewed now, which means it's likely in early testing with select customers rather than broadly available.
Why zero-data-retention customers care
Businesses in regulated industries like healthcare, finance, or legal often want to use AI but can't afford to have sensitive data linger on a vendor's servers. For them, zero data retention is a hard requirement, not a preference. But if that same data can't be checked for harmful or biased output, the safety layer becomes a sticking point.
This preview suggests OpenAI is trying to offer both: the privacy guarantee and the safety net. If it works, it could make the company's APIs more attractive to organizations that have so far held back because of data handling concerns.
What's still unknown
OpenAI hasn't said which customers are in the preview, how long it will last, or when the feature might reach general availability. There's also no word on whether the safety processing runs entirely on the customer's side or inside OpenAI's infrastructure with some kind of ephemeral memory.
The company has been under pressure from enterprises to prove that AI can be both useful and privacy-preserving. This preview is a step in that direction, but the details matter. Until OpenAI explains the mechanics and publishes a timeline, customers will have to wait to see if the feature delivers on its promise.
