AI-generated: This article was created and published automatically by LinkLoot and was not substantively reviewed by a human editor.

OpenAI keeps frontier-model ZDR with a private safety layer

OpenAI says eligible API customers will continue to receive Zero Data Retention access to its frontier models. The company is testing Private Safety Processing, an architecture intended to detect dangerous patterns across multiple interactions without giving OpenAI personnel access to the underlying prompts or responses.

The preview matters for organizations that handle proprietary code, financial records, health information, or regulated data. More capable agents need broader behavioral context for safety monitoring, but retaining that context can conflict with enterprise security and privacy requirements.

Key takeaways

• OpenAI says eligible API customers will retain its existing ZDR commitment for frontier models.
• Private Safety Processing is being tested with early customers and is not yet a generally available control.
• Content can remain on customer-controlled infrastructure; OpenAI is also developing storage encrypted with customer-controlled keys.
• A broader rollout and technical white paper are planned for September 2026.

OpenAI's ZDR commitment

Under OpenAI's description of Zero Data Retention, prompts and model responses are not retained after a request is processed, customer content is unavailable to OpenAI personnel for review, and enterprise data is not used for model training unless the customer opts in.

The commitment applies to eligible API customers. Axios reports that Private Safety Processing is not designed for consumer ChatGPT subscriptions, and the existing data controls for Free, Go, Plus, and Pro users remain unchanged.

OpenAI also identifies a legal exception: images flagged as possible child sexual abuse material may still be retained for manual review and reporting, including in ZDR deployments.

How Private Safety Processing separates content from signals

Existing ZDR-compatible protections inspect interactions individually. Private Safety Processing extends that analysis across related interactions, where repeated safeguard probing, coordinated activity, or an agent continuing after a stop instruction may reveal risks that one request does not.

For customer-hosted deployments, the content remains on infrastructure controlled by the customer. OpenAI is also developing an alternative where content sits on its infrastructure but is encrypted with keys controlled by the customer. OpenAI says its personnel would not possess those keys.

When the automated system identifies a risk, OpenAI receives a narrowly defined signal describing the activity type rather than the underlying conversation. Customers retain the information needed to investigate alerts and may voluntarily share relevant material when appealing an enforcement decision or assisting an abuse investigation.

What changes for enterprise buyers

The announcement gives security and compliance teams a reason to keep ZDR-compatible frontier models in their architecture plans instead of assuming stronger models will always require provider-readable logs. It does not remove the need to check endpoint-specific retention terms, feature persistence, regional processing, contractual exceptions, and the handling of files, vector stores, or other stateful services.

Private Safety Processing remains an early customer test. Until OpenAI publishes the planned technical paper, buyers cannot independently assess its threat model, cryptographic design, signal precision, false-positive rate, or operational response process.

The September milestone

OpenAI plans to begin rolling out Private Safety Processing and publish a technical white paper in September. That document should determine whether the preview becomes a practical compliance control or remains a vendor commitment that customers must evaluate mainly through contracts and architecture reviews.

Source check

• OpenAI's announcement defines the ZDR commitment, storage designs, limited safety signals, legal exception, and September plans.
• Axios reporting independently confirms the early-customer test and clarifies that consumer ChatGPT plans are not covered by this system.