OpenAI clarifies Private Safety Processing: automated abuse detection across conversations without retaining data

OpenAI clarified how Private Safety Processing works, a system running in preview for selected API customers. According to the company, the system automatically detects misuse of AI models across multiple related conversations at once, rather than only within a single session as the existing Zero Data Retention (ZDR) does. Monitoring takes place without human review of conversation content — if the system detects suspicious activity, it sends OpenAI only a narrowly defined signal about the type and severity of the activity, without giving the company access to the actual inputs or outputs. According to the company, this also makes it possible to detect distributed attempts at misuse, where a user splits harmful requests (such as malware development) across multiple sessions to evade detection within a single conversation. If the signal indicates that intervention is needed, OpenAI contacts the customer to request additional context; the customer can then share data at their own discretion. According to OpenAI, a technical white paper on the system is due to be released in September.

OpenAI also reaffirmed that the Zero Data Retention policy remains in effect for eligible API customers using frontier models. ZDR operates through automated agents that monitor abuse within a single session without retaining customer data; Private Safety Processing extends this principle to analysis across multiple conversations. According to the company, customer data remains on their own infrastructure or is stored in encrypted form, with the customer holding the keys. Aleah Houze, head of product policy at OpenAI, said that risks often become apparent only over the course of multiple conversations.

The approach taken by OpenAI contrasts with the policy at Anthropic, which retains customer data for 30 days for safety review for so-called “covered models” — including models in the Mythos series and future models with similar capabilities. This policy was announced in July and raised concerns among some customers working with sensitive data. Anthropic also states that human review of data is possible only through controlled access involving a small number of approved reviewers and that each such session is recorded in an immutable log; for other models, Anthropic largely adheres to ZDR.