OpenAI tests safety system that keeps customer data private

OpenAI began testing Private Safety Processing, a monitoring system designed to flag risky behaviour that spans multiple conversations without giving OpenAI staff access to the underlying prompts or responses. Microsoft and Databricks are early testers of the system, which OpenAI says sends only a narrowly defined safety signal rather than exposing customer content. A broader release and a technical paper are expected in September.

Under OpenAI's zero-data-retention deployments, customer content stays on infrastructure the customer controls, and content stored with OpenAI-provided storage is encrypted with keys the customer holds, so OpenAI personnel cannot access it directly. The approach contrasts with Anthropic's policy, announced in July, which permits retaining sessions from certain systems for up to 30 days for safety review.

Building safety monitoring around signals rather than content is a bet that pattern detection can work without the underlying transcripts, an approach enterprise customers handling sensitive data are likely to weigh against rivals' retention policies.