OpenAI to detect risks across chats without seeing content

The background is that as AI models take on longer, more complex tasks, serious safety risks may only become visible across multiple interactions. Existing safety systems evaluate each interaction individually, missing patterns that span conversations. This is why OpenAI is now previewing Private Safety Processing, which extends those protections across related interactions without exposing content to personnel.

For organizations handling sensitive data, the challenge is balancing safety with privacy. Recent frontier-model deployments forced some customers to allow retention for monitoring, which conflicts with their security commitments. Private Safety Processing aims to solve that by keeping content encrypted with customer-controlled keys, so OpenAI personnel cannot access it even when flagged.

The stakes are high. OpenAI must convince customers that the automated signals are precise enough to catch real abuse without false flags. The test is whether this approach provides genuine safety while preserving the trust that ZDR promises. More details are expected in September, when the white paper is released.