OpenAI Introduces Private Safety Processing for Zero Data Retention

OpenAI unveiled on August 19, 2026 a new security system called Private Safety Processing, designed to detect abuse patterns in frontier models without requiring the company to access customer data. The technology extends existing Zero Data Retention (ZDR) offerings with automated safeguards that identify risks across multiple interactions, while inputs and outputs are deleted after processing.

What is Zero Data Retention?

Zero Data Retention is a contractual configuration for Large Language Model (LLM) APIs where providers do not permanently store customer inputs and outputs after returning responses. The commitment applies to stateless API interfaces and is available exclusively for approved enterprise API endpoints – not for ChatGPT browser sessions, team plans, or consumer accounts.

OpenAI's ZDR policy guarantees that prompts and model responses are not stored after request processing and customer content is not available to OpenAI personnel for review. Since March 1, 2023, data sent to the OpenAI API is no longer used for training or improving models unless customers explicitly consent to share data.

Default Data Storage and Exceptions

By default, OpenAI creates abuse logs for all API usage, stored for up to 30 days unless longer retention periods are legally required or needed to protect services or third parties. Eligible customers can exclude their content from these abuse logs by obtaining approval for Zero Data Retention or Modified Abuse Monitoring.

These controls require prior approval from OpenAI and acceptance of additional requirements. Approved customers assume responsibility for ensuring that users comply with OpenAI's policies and fulfill moderation and reporting obligations under applicable law.

Private Safety Processing: Automated Pattern Detection Without Data Access

The most serious AI safety risks are not always visible in individual interactions. Potentially harmful intent often only becomes apparent when multiple interactions are examined together. Similar risks arise when malicious actors repeatedly test safeguards, coordinate activities across accounts, or disguise threats as routine research.

Aleah Houze, Head of Product Policy at OpenAI, explained on August 19, 2026: "We see with more powerful frontier models that risks often aren't detectable from a single prompt-response pair, but when you look at multiple interactions over time."

Private Safety Processing builds on automated safeguards already used in ZDR deployments. While existing ZDR-compatible security systems evaluate interactions individually, Private Safety Processing extends protection across related interactions. Automated systems can identify potential abuse patterns and return limited security signals without making underlying prompts or responses accessible to OpenAI personnel.

The system leverages customer content regardless of location – whether on customer-controlled infrastructure or in OpenAI-provided storage with customer-controlled encryption keys.

Distinguishing Between No-Training and Zero Data Retention

A critical distinction exists between two separate commitments: a no-training policy means providers do not use data to improve models, while ZDR means inputs and outputs are deleted after processing. The existence of one commitment does not automatically guarantee the other.

Anthropic Announces Mandatory 30-Day Storage

On August 19, 2026, Anthropic announced it will implement mandatory 30-day data retention for its most powerful models. According to reports, the decision is viewed as unpopular with customers expecting zero retention – a contrast to OpenAI's approach of expanding ZDR capabilities.

Business Context and Implementation

Zero Data Retention comes with higher pricing compared to standard offerings. The primary benefit lies in the ability to send sensitive data to frontier models without creating the audit trail problem that public cloud services entail. The provider commits in writing that data will not be stored beyond inference, will not be logged for training, and in many configurations will not even be used for abuse monitoring.

ZDR applies only to approved enterprise endpoints. When ZDR is enabled, the store parameter for the /v1/responses and /v1/chat/completions endpoints is always treated as false, even if requests attempt to set the value to true. Some endpoints and features listed as non-ZDR-compatible may still store application state even with ZDR enabled.

Customers interested in ZDR should contact OpenAI's sales team to learn more about offerings and clarify eligibility. Organizations can also enter into a Business Associate Agreement (BAA) with OpenAI for relevant use cases.