Offering Zero Data Retention for frontier models

OpenAI has announced the preview of Private Safety Processing, a new capability designed to allow the company to maintain its Zero Data Retention promise for eligible API customers as frontier models become more capable. According to OpenAI, existing safety systems evaluate each interaction individually, but complex tasks and advanced risks often only become visible across multiple interactions. Private Safety Processing extends automated protections to identify patterns across related interactions without exposing underlying prompts or responses to OpenAI personnel. For Zero Data Retention deployments, customer content remains on infrastructure controlled by the customer. OpenAI is also developing an option where content is stored on OpenAI infrastructure but encrypted with customer-controlled keys, ensuring personnel cannot access the underlying data. When automated systems identify potential misuse, OpenAI receives a narrowly defined signal indicating the activity type, allowing for enforcement decisions without human review of the content. Customers retain control over their information, which is critical for organizations handling sensitive financial, health, and proprietary data that must meet strict security obligations. OpenAI began rolling out Private Safety Processing to API customers in phases starting in September 2026, alongside plans to release a technical white paper. The initiative is being shaped through collaboration with customers across various industries, regions, and company sizes to balance advanced AI safety with rigorous data privacy.