Offering Zero Data Retention for frontier models

What changed

Source summary

OpenAI has extended automated safety monitoring to work with Zero Data Retention API deployments via a new "Private Safety Processing" system, which flags potentially harmful usage patterns without OpenAI staff accessing prompts or retaining customer content. It is in preview for eligible API customers now, with a wider rollout planned for September.

Why it matters to your work

Practical interpretation

If your organisation handles sensitive client data and has been avoiding OpenAI's API over privacy concerns, this closes a gap — you can now get abuse-monitoring safeguards without your data being stored or reviewed by OpenAI staff. Relevant mainly to organisations using the API under a Zero Data Retention agreement (an enterprise-level setup), not everyday ChatGPT users.

When this could help

Relevant now for procurement and compliance teams evaluating OpenAI's data handling commitments, especially for regulated industries (health, finance, education). No immediate user action required — ZDR is a contractual/API-level feature.

Zero Data Retention (ZDR) means OpenAI does not retain prompts or responses after processing and does not use them for training. Private Safety Processing (preview, rolling out from Sep 22 2026) adds automated pattern detection without staff content access — only a narrow signal on risk type is shared. Customer-key encrypted storage is still in development.

What to check

• No specific models or API endpoints are named as eligible for ZDR. Customer-key encrypted storage is described as still in development with no timeline. Source ↗
• Private Safety Processing preview means automated safety checks still occur — ZDR does not mean zero processing. The distinction matters for compliance assessments. Source ↗