Sunday, Aug 23 | --:--
Back to home

OpenAI Keeps Zero Data Retention on Frontier Models and Previews Private Safety Processing

On August 19, 2026, OpenAI said it will keep Zero Data Retention for eligible frontier-model API customers and previewed Private Safety Processing: automated misuse detection across related interactions that returns a narrow safety signal without exposing prompts or responses to OpenAI staff. Early testers include Microsoft and Databricks; broader rollout and a white paper are planned for September.

Tech Insights Reporter 5 min read San Francisco, CA
Cover illustration for OpenAI Keeps Zero Data Retention on Frontier Models and Previews Private Safety Processing

TLDR

OpenAI on Wednesday, August 19, 2026 published Offering Zero Data Retention for frontier models. Eligible API customers keep the ZDR promise: prompts and responses are not retained after a request; personnel cannot review customer content; enterprise data is not used for training unless the customer opts in. The new piece is Private Safety Processing—a preview of cross-session automated monitoring that is meant to work with ZDR, not replace it. Consumer ChatGPT Free / Plus / Go / Pro settings are unchanged.

What shipped

Item Detail (OpenAI Company/Safety primary + Axios/TC/Bloomberg)
Date August 19, 2026
ZDR Eligible API / enterprise frontier deployments; no retained prompts or completions
Private Safety Processing Automated pattern detection across related interactions; OpenAI staff see a narrow risk signal, not the underlying text
Where content lives Customer-controlled infrastructure (ZDR) or OpenAI storage encrypted under customer-held keys (in development)
Early testers Microsoft and Databricks (Bloomberg via Moneycontrol)
Rollout Broader launch + technical white paper planned September 2026
Spokesperson Aleah Houze, Head of Product Policy: risks now show up over time, not in a single prompt pair
Exception Images flagged as potential CSAM can still be retained for legally required review even under ZDR

Existing ZDR-compatible monitors scored one request at a time. PSP is OpenAI’s answer to agentic, multi-turn abuse without copying Anthropic’s 30-day retention on covered frontier models (Fable 5 / Mythos 5)—the contrast Axios and TechCrunch drew the same day.

Limits: architecture is not public until the September paper. Consumer ChatGPT is out of scope. ZDR eligibility is unchanged in kind; the news is the frontier + PSP pairing.

Product-line de-dupe: not a dual-date of Aug 7 Astra or Aug 18 pacing. This is enterprise data-governance GTM, not a training-calendar pause.

Why this story matters

Enterprises have been told they can have frontier models or zero retention, not both, once safety teams want a 30-day window. OpenAI is betting it can sell Sol-class API into regulated buyers by putting the monitor on the customer’s side of the key. Watch: whether September’s paper actually describes a system that catches Hugging Face-style agent chains without a human ever reading the transcript.

Sources

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading