OpenAI Previews Private Safety Processing to Widen ZDR

TechCrunch reported on 19 August 2026 that OpenAI is previewing Private Safety Processing for select customers. The automated system watches for abuse across multiple conversations while retaining none of the customer’s data, a design OpenAI says widens Zero Data Retention.

What Private Safety Processing is supposed to do

Zero Data Retention, or ZDR, is the enterprise promise that customer content does not sit on the lab’s disk after the call. The usual tradeoff is that abuse checks then have to happen in the moment, on one session. Private Safety Processing is OpenAI’s attempt to keep the no-retention claim and still watch a longer arc.

The company describes it as long-horizon safety monitoring. An automated system looks at inputs and outputs across more than one conversation. If it triggers, it sends a narrowly defined signal. OpenAI may then reach out. The customer may share data at its discretion. That last step is the privacy hinge: the lab does not keep the corpus up front, and any later share is optional.

TechCrunch frames the preview as a contrast with Anthropic’s recent retention posture on some models. That rivalry is context. The product fact is OpenAI’s: select customers, preview, no retained customer data, multi-conversation watch, narrow signal, optional follow-up share.

What a trigger does, and what it does not

A narrowly defined signal is not a transcript dump. OpenAI may decide whether it needs to talk to the customer. The customer may then choose to hand over more. If they refuse, the facts used here do not say what happens next. Do not invent an enforcement ladder.

This is still a preview for select customers. It is not a default on every API key. A skill that tells a client “OpenAI never sees anything, ever” is already too clean. ZDR plus a long-horizon watcher is a different sentence: the lab does not retain the data, and it still runs an automated abuse pass across sessions.

Write that sentence into enterprise runbooks. If your security team asked for ZDR because it did not want the lab to keep a store, this preview is aimed at that ask. If your security team asked for a human-reviewed archive, this is not that product.

How to write the skill around a signal, not a store

Keep tool-use logs on your side if you need an audit trail. Private Safety Processing is the vendor’s monitor, not your compliance archive. A skill that performs regulated work should still emit model_used, a redaction rule, and a stop when the user asks for something the policy forbids.

OpenAI has also slowed Astra training when it could not rule out a Critical cyber rating, and it has reshuffled preparedness. Those stories are about how the lab governs models. This story is about how it watches customers without keeping their text. Cite them as neighbors, not as one incident.

Do not promise a customer that OpenAI cannot see a long-running abuse pattern. Promise that, in this preview, the lab says it keeps none of the customer’s data unless the customer later shares it. The citation sentence is OpenAI, Private Safety Processing, 19 August 2026, TechCrunch. Preview for select customers. Automated abuse watch across multiple conversations. Retains none of the customer’s data. Widens ZDR. Long-horizon safety monitoring. Narrowly defined signal if triggered. OpenAI may reach out. Customer may share data at discretion.

Sources

• OpenAI seeks to one-up Anthropic with new customer privacy protections — TechCrunch, 19 August 2026