OpenAI is previewing Private Safety Processing, a system designed to identify safety risks across multiple related API interactions without retaining customer prompts or model responses. This update supports the company's Zero Data Retention (ZDR) promise by allowing automated systems to detect patterns in potentially harmful intentions while ensuring OpenAI personnel never access the underlying content.
- The system evaluates patterns across related interactions rather than evaluating each interaction individually, addressing risks that emerge over complex tasks or from coordinated probing.
- Customer content remains on infrastructure controlled by the customer or is stored on OpenAI infrastructure encrypted with keys controlled by the customer.
- When a risk is identified, OpenAI receives only a narrowly defined signal indicating the type of activity, without access to the flagged content.
- The feature is currently being tested with early customers and is planned for rollout in September alongside a technical white paper.
This approach aims to resolve conflicts between safety monitoring requirements and enterprise security obligations, ensuring that sensitive data like financial records or health data remains under customer control while maintaining effective safeguards.