OpenAI adds an AI safety layer to detect misuse without retaining enterprise data

OpenAI has introduced a new safety feature called Private Safety Processing that enables enterprises to detect misuse of its AI systems across multiple interactions without retaining any prompts or responses. This capability is designed to identify patterns without giving OpenAI personnel access to the underlying content, preserving its Zero Data Retention (ZDR) commitments.

OpenAI explained that the system correlates activity across related interactions rather than analyzing each prompt in isolation, helping to detect risks that may only become apparent when multiple interactions are viewed together. It is currently being tested with eligible enterprise and API customers. By operating on encrypted data within enterprise-controlled infrastructure or customer-controlled encryption keys, Private Safety Processing can generate a narrowly defined signal indicating the type of activity involved without exposing the underlying prompts or responses to OpenAI personnel.

This approach addresses a gap in existing safety controls, as harmful intent may only become clear when multiple interactions are taken into account. The introduction of Private Safety Processing highlights differing safety approaches among AI providers. While OpenAI focuses on detecting misuse patterns across interactions while preserving zero data retention, some providers retain customer interaction data for monitoring purposes, reflecting a different approach to identifying risks spanning multiple requests.