OpenAI is previewing a new service for select customers called Private Safety Processing. The company presents this as a privacy-centric expansion of its existing Zero Data Retention (ZDR) approach: automated agents monitor for potential abuse across multiple conversations while the system itself does not retain customer data.
What the technology does and how it works
Private Safety Processing performs long-horizon safety monitoring by correlating inputs and outputs across multiple sessions rather than inspecting single sessions in isolation. When an automated agent detects suspicious patterns, the system can generate a narrowly defined signal that alerts OpenAI to a specific type of activity.
If such a signal occurs, OpenAI will determine whether enforcement or further action is necessary. If so, the company may contact the customer for additional context or to work with them on the issue. According to OpenAI, the customer retains control and can choose whether to share more data with the company.
Why this matters
The feature is designed to catch malicious uses of AI that unfold over time and across sessions. For example, a bad actor trying to engineer malware or plan a cyberattack might break their requests into many small queries to avoid detection. Private Safety Processing can link those separate interactions to surface signs of abuse without initial human review of the user’s conversations.
How this contrasts with Anthropic’s policy
The announcement stands in contrast to Anthropic’s recently announced data-retention policy. In July, Anthropic said it would retain user session data for 30 days for certain “covered models,” which include models it classifies as Mythos and “future models with similar capabilities.” That policy was justified on safety grounds to allow the company to analyze potential impropriety, but it has worried some enterprises that handle sensitive data and prefer their data not be held or inspected by the AI lab.
Anthropic notes that human review of customer data can occur under its policy, but only through a controlled access path involving a small set of approved reviewers, and that each review session is recorded in a tamper-proof log.
Relation to Zero Data Retention
OpenAI and many other AI companies already use a Zero Data Retention model in which API-level agents scan for abuse on a per-session basis without storing customer data long term. Private Safety Processing expands that model by enabling automated correlation of multi-session interactions, while maintaining that the monitoring is automated and that data is not retained by default.
Competitive context
The move comes amid heightened competition between OpenAI and Anthropic. A recent report indicated OpenAI’s Q2 growth was slower than Anthropic’s. Anthropic’s annualized revenue run rate is reportedly $65 billion, and some investors have suggested the company could pursue an IPO valuing it as high as $2 trillion; OpenAI is also working on its own IPO plans.
Summary
Private Safety Processing aims to balance enterprise privacy with effective abuse detection by identifying multi-session misuse without long-term retention of customer conversations. The service will be previewed to a limited set of customers, and OpenAI says it will contact customers only if a narrowly defined automated signal indicates potential misuse.
