OpenAI previews privacy-focused system for detecting AI misuse

OpenAI is previewing Private Safety Processing with early customers seeking greater certainty about how their data will be protected as AI systems become more capable. The system identifies patterns across related interactions while restricting OpenAI personnel from accessing the underlying content. The company plans to start rolling it out and publish a technical white paper in September.

“No AI lab can address emerging risks alone. Private Safety Processing reflects that approach and is being shaped by customers across industries, regions, and company sizes,” OpenAI wrote.

For eligible API customers using Zero Data Retention (ZDR), prompts and model responses are not retained after a request is processed. An exception applies to images flagged as potential CSAM, which may be retained for manual review and reporting. OpenAI says enterprise customer data is not used to train its models unless customers opt in.

Content Access Controls (Source: OpenAI)

ZDR deployments keep content on infrastructure controlled by the customer. OpenAI is developing another option that would store it on the company’s infrastructure using customer-controlled encryption keys. In both configurations, automated systems can identify potential misuse and return limited safety signals without revealing prompts or responses.

Safety analysis across interactions

Private Safety Processing builds on automated protections used in ZDR and other deployments. Existing safeguards evaluate requests individually, while the new system analyzes related activity to detect patterns of potential misuse.

“In healthcare, protecting massive amounts of sensitive data and ensuring its accuracy and integrity are fundamental to earning the trust of clinicians and patients. Having the chance to work directly with OpenAI’s product, engineering, policy, and leadership teams to help shape those practices has made for an unparalleled partnership. That level of collaboration on trust and security is uncommon. They listen, they take action, and that gives us confidence in OpenAI. That level of partnership on trust and security is just not common,” said Zach Powers, CISO, Abridge.

The system can analyze data stored on customer-controlled infrastructure or through OpenAI’s planned encrypted storage option. When it detects a potential risk, the company receives a defined signal indicating the type of activity involved. That information can inform enforcement decisions.

Customers can investigate alerts using information available in their own systems and share relevant details with OpenAI to appeal a decision, clarify legitimate activity or support an investigation into verified abuse.