OpenAI has previewed Private Safety Processing, a new system designed to spot patterns of misuse across multiple interactions while remaining compatible with its Zero Data Retention service for API customers.
Zero Data Retention currently guarantees that OpenAI does not retain eligible customers' prompts or responses after a request is processed, and does not use enterprise data for training without explicit opt-in. One exception applies regardless of ZDR status: content flagged for potential child sexual abuse material is retained for manual review and reporting, as legally required. Existing ZDR-compatible safety systems otherwise assess each interaction in isolation, which OpenAI says limits their ability to catch risks that only become visible across a sequence of interactions, such as repeated probing of safeguards, coordinated abuse across accounts, or an AI agent that continues acting after being told to stop.
Under the new system, customer content stays either on infrastructure the customer controls or, in an option OpenAI is also developing, on OpenAI's own infrastructure but encrypted with keys held by the customer. Automated systems analyse patterns across interactions and can flag a narrowly defined signal, such as the type of activity involved, without giving OpenAI staff access to the underlying content. Customers can then investigate flagged activity themselves, and may choose to share content with OpenAI if they want to appeal a decision or support an abuse investigation.
Private Safety Processing is currently being tested with early customers, with OpenAI citing feedback from organisations including Glean, Databricks, Abridge and Microsoft. The company plans to begin a wider rollout, alongside a technical white paper, in September.
