OpenAI Updates ZDR Safety Signals

OpenAI has released a Private Safety Processing preview that ensures safety for API customers in Zero Data Retention (ZDR) environments, where prompts and responses are not stored after processing.

This feature is designed to detect patterns across relevant interactions without OpenAI personnel accessing the actual content. When risks are detected, only limited safety signals are transmitted to OpenAI, while customer-controlled infrastructure and customer-held encryption keys remain unchanged from existing models.

Auditability Issues

This architecture, which prevents providers from reading content, raises questions about auditability. For customers to trust the system, the following elements are required:

• Which policies were triggered
• The scope of the interaction window evaluated
• False positive appeal procedures
• The scope of evidence remaining in customer systems

Proposed measures to enhance trust include public signal schemas, reproducible customer-side notifications, independent audits, and cryptographic attestation.