OpenAI previewed Private Safety Processing with zero-data-retention compatibility; early customers include Microsoft and Databricks, with a September rollout planned.

OpenAI previewed Private Safety Processing, a system intended to strengthen safeguards while remaining compatible with zero data retention. It is designed to detect risk patterns spread across multiple interactions rather than examining only an isolated prompt and response.

OpenAI said longer and more autonomous AI workflows can produce risks through combinations of interactions. The system is intended to protect against attacks by malicious users and misaligned AI agents without giving OpenAI personnel access to the underlying raw customer content.

The architecture does not retain customer prompts or model responses, and OpenAI personnel cannot view them. Enterprise customer data will not be used for model training unless the customer explicitly opts in.

Microsoft and Databricks were identified as early customers participating in testing. OpenAI expects a broader rollout in September and plans to publish a technical white paper alongside the release.