OpenAI has announced Private Safety Processing (PSP), a system designed to resolve one of enterprise AI's most persistent tensions: how to monitor complex AI interactions for safety risks without actually looking at the data.

How Private Safety Processing Works

PSP is built to uphold OpenAI's Zero Data Retention commitments for API and enterprise clients. Under those commitments, no customer data is stored, reviewed, or used for training. The system transmits only limited safety signals to OpenAI, keeping the actual content of interactions completely opaque to the company. In effect, PSP extends safety monitoring into a regime where OpenAI otherwise retains nothing to review.

The need for such a system has grown as AI deployments have become more sophisticated, increasingly involving longer sequences of interactions rather than simple one-off queries. A single exchange may look perfectly benign, but string twenty together and the pattern could reveal someone systematically probing safeguards, or an AI agent drifting off-mission. PSP is designed to catch those multi-turn patterns without compromising the privacy guarantees that enterprise customers demand.

Customers have two options for handling their data: keep it entirely within their own infrastructure, or store it encrypted on OpenAI's servers with encryption keys that the customer controls.

OpenAI announced the system on August 19, 2026, with a phased rollout to API customers beginning on September 22, 2026. A technical white paper accompanied the broader availability, laying out the architectural details for security teams to evaluate — a reference point as the rollout begins and enterprises weigh the architecture against the privacy guarantees their own deployments require.

The Enterprise Privacy Arms Race

OpenAI did not build PSP in a vacuum. Anthropic established a 30-day data retention requirement starting in June 2026 for certain frontier models. OpenAI is positioning PSP as the opposite approach: rather than requiring retention, the system is engineered to function while retaining nothing. For enterprise buyers, the divergence outlines two distinct models for reconciling safety oversight with data control: retention-based review on one side, retention-free processing on the other.

Early development partners helped shape the feature before its public release. Microsoft, Databricks, Glean, and Abridge all provided feedback during the development process.

What It Does Not Cover

One important caveat remains: Zero Data Retention features are unavailable for consumer tiers of ChatGPT. That includes the Free, Plus, Go, and Pro plans. PSP is exclusively an enterprise and API offering, drawing a clear boundary around where multi-turn safety monitoring of this kind is available.