# Zero Retention, Yet Risks Still Caught! OpenAI Previews 'Private Safety Processing'

Hi everyone, it's Shii-chan! Today I've got news about OpenAI's approach to privacy and safety!

## What was announced?
OpenAI News reaffirmed its commitment to "Zero Data Retention" (ZDR) for eligible API customers. ZDR is a promise that OpenAI does not retain prompts or model responses once a request has been processed. Customer content isn't even available to OpenAI personnel for review, and it's not used to train models unless a customer explicitly opts in.

Alongside that, OpenAI previewed a new system called "Private Safety Processing," designed to identify patterns across multiple interactions without exposing the underlying content to OpenAI personnel.

## The story so far
Existing ZDR-compatible safety systems evaluated each interaction individually. But as models take on longer, more complex tasks, serious risks increasingly only become visible when you look across multiple interactions together — not in a single one. Bad actors might repeatedly probe safeguards, coordinate across accounts, or disguise harmful activity as legitimate research. Agentic tasks add another wrinkle: a system could become misaligned with a user's intent by continuing to act even after being told to stop.

Some recent frontier-model deployments required customers to let their AI provider retain sensitive content for safety monitoring — a requirement that conflicted with the security obligations or commitments of many organizations.

## What changes
Private Safety Processing extends the existing ZDR protections across related interactions instead of evaluating each one in isolation. For ZDR deployments, content stays on infrastructure the customer controls. OpenAI is also developing an option to store content on OpenAI's own infrastructure, encrypted with keys the customer controls — OpenAI personnel don't hold a copy of those keys, so they can't access the content.

In both cases, automated systems identify potential misuse and return only a narrowly scoped safety signal, without exposing the underlying prompts or responses. When a risk is flagged, OpenAI only receives a signal indicating the type of activity involved, similar to its existing safety systems, and uses that to decide whether enforcement is needed — OpenAI personnel still don't get access to the actual content, even when it's flagged.

Customers can investigate alerts using information in their own systems, and can choose to voluntarily share information with OpenAI if they want to appeal a decision, clarify legitimate activity, or support an abuse investigation.

## Dive Deep
Private Safety Processing is currently being tested with early customers. The announcement includes a quote from Sunil Agrawal, Chief Information Security Officer at Glean: "Enterprise AI adoption depends solely on customer control of data, with no direct or derivative use beyond the chosen service. OpenAI's no-training commitment and ZDR give Glean confidence to build with OpenAI."

OpenAI plans to begin rolling out Private Safety Processing, along with a technical white paper, in September 2026.

## Wrap-up
- OpenAI reaffirmed its Zero Data Retention (ZDR) commitment
- Previewed a new system called "Private Safety Processing"
- Unlike older ZDR safety systems that evaluated interactions individually, the new approach can spot patterns across related interactions
- Content stays on customer-controlled infrastructure, or on OpenAI infrastructure encrypted with customer-controlled keys that OpenAI personnel can't access
- Full rollout and a technical white paper are planned for September 2026
- Highly relevant for anyone at a company handling sensitive data who also cares about AI safety guarantees!