OpenAI has reaffirmed its commitment to Zero Data Retention (ZDR) for eligible API customers using frontier models while introducing the new Private Safety Processing.
This safety architecture is designed to detect multi-session misuse without exposing the underlying prompts or responses to OpenAI personnel.
Announced on August 19, 2026, this initiative addresses a critical challenge in enterprise AI security: maintaining effective abuse detection for increasingly autonomous systems while allowing organizations to retain control over highly sensitive data.
OpenAI Zero Data Retention
ZDR means that after processing a covered API request, OpenAI does not retain the customer’s prompts or model responses. Additionally, customer content is not accessible for personnel review, and enterprise API data is not used for model training unless the customer explicitly opts in.
This is particularly important for organizations handling regulated and confidential material, such as health records, financial data, proprietary research, legal communications, and internal business plans, since retaining prompt logs can raise concerns about privacy, compliance, breach notification, and insider risks.
Traditional ZDR-compatible protections operate on a per-interaction basis, potentially overlooking evolving threats. Examples include iterative guardrail probing, coordinated malicious activity across related accounts, or an AI agent gradually exceeding a user’s intended authority.
Private Safety Processing aims to introduce cross-interaction detection while preserving the privacy boundaries that make ZDR beneficial for sensitive API use cases.
OpenAI’s design features two data-handling models. In ZDR deployments, customer content remains on infrastructure controlled by the customer.
| Feature | Existing ZDR Safety Controls | Private Safety Processing |
|---|---|---|
| Analysis scope | Individual request or interaction | Patterns across related interactions |
| Prompt/response retention | Not retained after processing for covered customers | Supports ZDR while enabling broader automated analysis |
| Human access | OpenAI personnel cannot review customer content | Personnel still do not receive underlying flagged content |
| Risk output | Per-request automated safety decisions | Narrowly defined signal describing potential risky activity |
| Customer-controlled deployment | Supported | Content may remain on customer infrastructure |
| OpenAI-hosted storage option | Not the focus of ZDR | Under development with customer-controlled encryption keys |
| Availability | Available to eligible API customers | Early-customer testing; broader rollout and technical paper planned for September |
Separately, OpenAI is working on hosted storage that will be encrypted with keys controlled by the customer, meaning OpenAI personnel will not have access to those keys.
Automated systems can assess associated activity and provide OpenAI with a limited risk signal, rather than readable prompts, outputs, or conversation history, to inform potential enforcement actions.
From a cybersecurity perspective, this approach is similar to privacy-preserving security telemetry: a service can receive alerts about categories, such as suspected policy violations, without automatically accessing the sensitive event details that triggered those alerts. Customers can continue to investigate alerts using their own logs.
They can voluntarily provide relevant information if they need to appeal an enforcement decision, explain legitimate research activities, or support an investigation into confirmed abuse.
This announcement is particularly relevant for security teams utilizing frontier models in areas such as agentic workflows, code analysis, incident response, vulnerability research, and proprietary data environments.
It has the potential to reduce the trade-off between data minimization and safety observability. However, customers should await the forthcoming technical white paper, which will provide details on implementation, scope boundaries, cryptographic assurances, eligibility criteria, and metadata handling limitations.
Currently, Private Safety Processing is being tested with early customers and is not yet broadly available.
Prevent incidents due to slow investigations. Power your Tier 1 with threat intelligence from 15K SOCs: Integrate TI Lookup in your SOC
 for eligible API customers using frontier models while introducing the new Private Safety Processing.){kind=link}