On August 19, 2026, OpenAI announced a system called "Private Safety Processing" that maintains Zero Data Retention (ZDR) for API customers while detecting abuse patterns across multiple interactions. The system is currently being tested with early adopters; a broader rollout and technical whitepaper are planned for September 2026.

Zero Data Retention means that the AI provider does not store customer inputs and model responses after processing. For companies with regulatory obligations or customer commitments regarding data processing, ZDR is often a compliance requirement. OpenAI guarantees eligible API customers that prompts and outputs are not retained after processing and are not available to OpenAI personnel for review.

Security Monitoring Without Data Access

Private Safety Processing uses automated systems to identify potential threats spanning multiple requests without granting OpenAI personnel access to the underlying content. Aleah Houze, Head of Product Policy at OpenAI, explained: "We see with more capable frontier models that risks often don't just emerge in single prompt-response pairs, but when you look at multiple interactions over time."

Houze gave as an example scenarios where someone asks about vulnerabilities in a company's software in one conversation and later asks about remote access or security tools – a pattern that could indicate a potential cyberattack. The system is designed to detect such patterns without OpenAI staff being able to see the actual prompts or responses.

Technical Implementation

Customer data remains either on customer-controlled infrastructure or on OpenAI infrastructure with customer-controlled encryption keys. In both cases, automated systems can identify potential abuse and return security signals. Classifier outputs and content flagged for violation review remain persistent on separate time horizons, even under ZDR.

The system works regardless of where customer data is stored. Existing ZDR-compatible security systems evaluate interactions individually; Private Safety Processing extends this to related interactions over time.

Risks in Frontier Models

OpenAI identifies several risks that are not visible in individual interactions:

  • Bad actors repeatedly testing safeguards
  • Coordination across multiple accounts
  • Threats disguised as routine research
  • Agentic tasks where a system diverges from user intent when it continues after the user has stopped it

According to OpenAI, some recent frontier model deployments required customers to have their AI provider retain sensitive content for security monitoring – which conflicts with the security obligations or assurances many organizations have to the people they serve.

Anthropic Requires 30-Day Data Retention

While OpenAI pursues zero-retention security, Anthropic announced an opposite strategy in August 2026. The company requires 30-day data retention for its most capable models – a decision Anthropic itself describes as "unpopular with customers accustomed to zero-retention" and which "poses real risks to our business success, especially if competitors don't follow."

According to its August 2026 risk report, Anthropic justifies this position by saying that 30-day retention is "essential to detect and prevent sophisticated attacks spanning multiple requests."

Scope and Limitations

Private Safety Processing is designed for eligible Enterprise and API customers – not for users of OpenAI's paid or subscription consumer ChatGPT plans (Free, Plus, Go, Pro). The existing privacy settings for these consumer users remain unchanged.

ZDR exclusively governs the storage of prompts and outputs on the provider side. Training-data opt-out is a separate protection that, according to OpenAI, applies by default. Enterprise customer data is not used for model training unless customers explicitly choose otherwise.

Stateful capabilities such as file storage, batch processing, and server-side agent states deliberately retain data and typically fall outside of the ZDR agreement or under separate feature-specific terms.

Connection to Other Security Initiatives

On August 18, 2026, OpenAI announced it would pause some work on training new models to address security concerns. OpenAI plans to publish technical reports of its findings in the coming weeks.