OpenAI previews cross-session safety monitoring without data retention
OpenAI is previewing Private Safety Processing for selected customers, describing an automated approach to safety monitoring that can assess inputs and outputs across multiple conversations while retaining none of the customer’s data. The service is intended to identify possible misuse that may be distributed over sessions rather than visible in a single exchange.
The announcement places customer privacy at the centre of an increasingly important enterprise AI question: how providers can detect harmful activity without storing or exposing sensitive prompts and outputs. OpenAI says the new capability expands the scope of its existing Zero Data Retention, or ZDR, approach.
Long-horizon monitoring without retained conversations
Under ZDR, agents working within the OpenAI API monitor sessions for potential abuse without retaining customer data. Private Safety Processing adds what OpenAI calls long-horizon safety monitoring. Instead of limiting assessment to one session, an agent can evaluate patterns across multiple conversations.
This matters because a malicious actor could divide requests across separate sessions in an effort to avoid detection. OpenAI gave the example of attempts to engineer malware for a cyberattack. When its system is triggered, it may transmit a narrowly defined signal identifying a specific type of activity, rather than provide OpenAI with the underlying conversation content for human review.
OpenAI says it can then determine whether enforcement is necessary. The company may contact the customer for context or to work on an issue, while the customer remains able to decide whether to share data. The design therefore separates automated detection from discretionary disclosure of customer content.
A contrast with Anthropic’s covered-model policy
The rollout arrives amid a visible competitive contest with Anthropic over enterprise safety and privacy. Anthropic’s policy for covered models permits retention of user sessions and their conversations for 30 days so the lab can analyze potential impropriety. The covered group includes Mythos-class models and future models with similar capabilities; Anthropic otherwise largely follows ZDR, except for those models, including Fable.
Anthropic says human review can occur only through a controlled access path involving a small set of approved reviewers. It also says every review session is recorded in a tamper-proof log that reviewers cannot suppress or modify. The policy has nevertheless concerned enterprises that process sensitive data and do not want it retained or inspected by an AI provider.
The policy debate also sits alongside the business pressure surrounding Anthropic disputes and Mythos business impact, where disputes involving Mythos have implications for Anthropic’s commercial position. OpenAI’s preview frames privacy-preserving monitoring as an alternative way to address misuse risks across longer patterns of activity.
What enterprise buyers should examine
For businesses using AI with sensitive information, the practical issue is not simply whether a provider offers safety monitoring. They should establish whether monitoring is confined to a session or can span sessions, whether content is retained, what data leaves the customer environment when an alert occurs, and under what conditions human review is possible.
Private Safety Processing remains a preview for selected customers, so its operational use will depend on its eventual availability and implementation. The business implication is clear: procurement teams should compare safety mechanisms and data-retention terms together, especially where multi-session misuse detection and confidentiality requirements must coexist.