Zero Data Retention for Frontier Models

OpenAI has released Private Safety Processing in preview, a new safety measure that allows API customers to maintain their Zero Data Retention (ZDR) policy, where prompts and model responses are not stored after request processing.

Existing ZDR-compatible safety systems had the limitation of evaluating each interaction individually. However, as AI models perform longer and more complex tasks, potential risks often only become apparent when viewing multiple interactions holistically. Repeated probing attempts by malicious actors, coordination across accounts, and intent drift during agentic tasks are difficult to capture with single-session analysis.

Private Safety Processing is designed to identify patterns across relevant interactions without OpenAI employees accessing customer content. Customer content either remains on infrastructure controlled by the customer or, if stored on OpenAI infrastructure, is encrypted with keys controlled by the customer. OpenAI employees do not possess these keys and therefore cannot access the original content.

When a risk is detected, OpenAI receives only limited safety signals indicating the type of activity. These signals are used to determine the need for enforcement actions, and customers can investigate alerts within their own systems and selectively share relevant information with OpenAI if necessary. The system is currently being tested with early customers and was released to provide predictability in content protection even as AI capabilities advance.