OpenAI is testing a new safety-monitoring system called Private Safety Processing that can flag patterns of misuse across multiple related conversations without giving the company’s own staff access to the underlying content, according to an OpenAI announcement published August 19, 2026.
The gap it closes
OpenAI already offers Zero Data Retention for some frontier-model deployments, meaning customer prompts and outputs aren’t stored once a request is processed. But OpenAI says the most serious safety risks — repeated probing for exploits, coordination across multiple accounts, or an AI agent quietly drifting off-task during a long-running job — often only become visible when reviewers can look across several interactions at once. Under strict zero-retention rules, that kind of cross-session review wasn’t possible, leaving some customers to choose between full monitoring and stronger privacy guarantees.
How it works
Private Safety Processing analyzes a customer’s related interactions using automated checks similar to those OpenAI already runs inside its zero-retention deployments, but extended across sessions instead of a single exchange. Customer content stays on the customer’s own infrastructure or is stored on OpenAI’s servers encrypted with keys the customer controls. When the system flags a risk, OpenAI says it receives only a narrowly defined signal describing the type of activity detected — not the prompts or responses themselves. Customers can then investigate flagged activity using their own logs, and can choose to share details with OpenAI if they want to appeal an enforcement decision.
Rollout timeline
OpenAI is currently testing the system with a small number of early customers. The company plans a broader rollout along with a technical white paper in September 2026 that should detail the underlying detection methods.
The move responds to a tension that has grown as AI agents take on longer, more autonomous tasks for businesses: the more independently a model acts, the more related interactions a safety system needs to see to catch problems, even as enterprise customers demand stricter data-handling guarantees. OpenAI has faced scrutiny of its own safety practices this year — it paused frontier-model training over a cybersecurity concern in a separate incident — while other labs, including Google with its open-sourced HEIR compiler for encrypted AI inference, have pursued their own approaches to reconciling privacy with oversight.