OpenAI has introduced new protocols for enterprise data privacy, specifically focusing on its Zero Data Retention (ZDR) policy. Under this policy, OpenAI commits to not retaining prompts or model responses after processing, ensuring that customer content remains inaccessible to personnel for internal review. This initiative is designed to address the privacy requirements of organizations handling sensitive financial, health, and proprietary data.
As AI models perform more complex and agentic tasks, identifying potential safety risks across multiple interactions becomes more difficult. To address this, OpenAI is previewing Private Safety Processing. This system allows automated safety tools to detect patterns across related interactions without requiring manual oversight or access to the underlying content. This maintains the privacy standards of ZDR while improving the ability to catch misuse.
For deployments, customer content remains on infrastructure the customer controls. OpenAI is also developing an encryption option where data is stored on OpenAI infrastructure but secured with customer-controlled keys. In this setup, OpenAI personnel cannot access the content, even when automated systems flag potential safety signals. This provides a path for organizations that require both high-level security and oversight capabilities.
This move responds to feedback from enterprise partners who need assurance that their data will not be used for model training or retained unnecessarily. Companies like Glean have noted that these commitments are critical for integrating AI into environments with strict regulatory or security obligations. OpenAI plans to roll out Private Safety Processing in September, supported by a technical white paper to help organizations understand the implementation details.
By moving away from manual content review and toward encrypted, automated safety processing, the company aims to balance the needs of AI safety with the demand for data sovereignty. This strategy aligns with broader enterprise goals of maintaining control over confidential business information while adopting modern language models.