tech
Offering Zero Data Retention for frontier models
Previewing Private Safety Processing, which strengthens safeguards across interactions while remaining compatible with ZDR.
TL;DR
- OpenAI introduces Private Safety Processing to strengthen AI safeguards for frontier models.
- The system is compatible with Zero Data Retention (ZDR), ensuring customers' prompts and responses are not retained by OpenAI.
- Private Safety Processing identifies potential misuse by analyzing patterns across related interactions, not just individual ones.
- Customer content remains on their infrastructure or is encrypted with customer-controlled keys on OpenAI infrastructure.
- Automated systems detect misuse and return safety signals without exposing underlying content to OpenAI personnel.
- This evolution is necessary as AI models handle longer, more complex tasks where risks may only be visible over multiple interactions.
- The system allows customers to investigate alerts and share information with OpenAI if needed for appeals or investigations.
- The feature is currently being tested and plans to roll out in September, with a technical white paper to follow.
- Images flagged for potential child sexual abuse material (CSAM) will continue to be retained for manual review and reporting, as legally required.