OpenAI Previews Cross-Session Safety Checks Without Keeping Customer Chats

OpenAI is previewing Private Safety Processing, a system for select customers that is meant to detect potential misuse across several conversations without retaining customer data. OpenAI says the monitoring is automated and does not send conversations to human reviewers as part of the routine process.

From one session to a longer pattern

The change is in the scope of the safety check. OpenAI’s existing Zero Data Retention approach monitors API activity for abuse on a session-by-session basis without retaining customer data. Private Safety Processing extends that approach by assessing inputs and outputs from multiple conversations over time.

That wider view addresses a specific evasion problem: activity that appears ordinary in isolation may form a concerning pattern when connected across sessions. OpenAI gave the example of a person spreading requests related to malware development across separate conversations to avoid detection.

A trigger starts a review process, not an automatic penalty

When the system triggers, OpenAI says it may receive a narrowly defined signal about a particular type of activity. The signal does not automatically result in enforcement: OpenAI may contact the customer for context before deciding whether action is necessary. If more investigation is needed, the customer may choose whether to share relevant data with OpenAI.

Anthropic’s policy applies to a defined model category

The comparison is narrower than a blanket divide over enterprise privacy. Anthropic largely follows Zero Data Retention outside covered models, according to the policy description. For covered models, however, its policy permits the company to keep sessions and their conversations for 30 days so it can analyze potential safety problems.

Anthropic says human review can occur through a controlled access path involving a small group of approved reviewers. It also says each review session is recorded in a tamper-proof log that reviewers cannot suppress or alter. That approach keeps material available for inspection under stated access controls; OpenAI’s preview instead puts its safety case on automated detection before any customer decides to provide more data.

The product claim is a new privacy-safety tradeoff

Private Safety Processing is still a preview limited to select customers. But its distinct promise is clear: OpenAI says it can look for misuse that unfolds across a longer sequence of interactions while preserving a no-retention boundary. The result is a different answer to the central enterprise question—how a provider can seek signs of misuse without holding the underlying conversations for later analysis.