OpenAI announced the preview of Private Safety Processing on August 19, 2026, to strengthen safety measures for its frontier models while maintaining Zero Data Retention. This update allows automated systems to identify patterns across related interactions without exposing customer content to OpenAI personnel. The initiative aims to address risks that may only become visible through multiple interactions, such as harmful intentions or misaligned behavior in complex tasks.
Private Safety Processing builds on existing automated protections in Zero Data Retention deployments. It enables systems to detect potential misuse by analyzing patterns without accessing retained customer content. Customer data remains on infrastructure the customer controls, or is stored on OpenAI infrastructure with encryption keys managed by the customer. OpenAI personnel do not have access to these keys, ensuring data privacy. When risks are identified, OpenAI receives limited safety signals to determine enforcement actions, without exposing the underlying content to its team.
The update follows feedback from enterprise customers who emphasized the need for predictability in data protection as AI systems grow more capable. OpenAI plans to roll out Private Safety Processing and release a technical white paper in September, with ongoing communication to ensure transparency and support for customers. The company reiterated its commitment to privacy and safety, highlighting collaboration with customers to build stronger safeguards.
Source: openai
