OpenAI announced an expansion of its Zero Data Retention (ZDR) capability with a new feature called Private Safety Processing, allowing enterprise customers to retain API data without OpenAI personnel having access to it. Under the system, customer content is stored either on customer-controlled infrastructure or on OpenAI servers encrypted with customer-managed keys—OpenAI holds no key copies. Automated safety systems can identify patterns and misuse signals across multiple interactions without exposing the underlying prompts or responses to OpenAI staff.
The motivation is clear: frontier models handling longer, more complex agentic tasks require safety monitoring across multiple interactions, but many enterprises face legal or compliance constraints against third-party data access. Private Safety Processing is currently in preview with early customers and represents an attempt to square the circle of safety and privacy. When a risk is flagged, OpenAI receives only a narrowly defined signal (similar to existing systems) without content visibility; customers can investigate using their own systems and optionally share details if appealing an enforcement decision.
For architects: this addresses a friction point in enterprise adoption of frontier models—compliance teams have historically blocked API deployment until data retention policies met internal standards. Customers cited include Glean, Databricks, Abridge, and Microsoft, signaling broad interest across fintech, insurance, and healthcare. As models scale to agentic deployments and multi-turn workflows, the ability to run safety monitoring without middle-layer data custody becomes table-stakes. This is a material shift in vendor trust economics for regulated industries.