On August 19, OpenAI posted a notice on its website aimed at API customers: eligible customers will continue to get Zero Data Retention (ZDR), meaning prompts and model responses aren't kept once a request has been processed. The same notice previewed a mechanism called "Private Safety Processing" — a way to spot signs of abuse across multiple related interactions without exposing the raw content to OpenAI employees.
Zero data retention isn't a new term. For customers in finance, healthcare, and legal, whether a vendor can promise "we don't store anything on our end" has long been a deal-breaker clause. Public reporting suggests this standalone notice was prompted by models taking on longer, more autonomous work: a single request may look harmless, but the risk is spread across a whole chain of sessions. Catching that kind of pattern requires keeping something to compare against — which runs headlong into zero retention.
Moving detection into an encrypted box
OpenAI's fix is to hand the comparable part back to the customer. Customer content can stay on infrastructure the customer controls, or be stored by OpenAI while the encryption keys stay with the customer. An automated system evaluates that encrypted content across sessions, looking at inputs and outputs from multiple conversations rather than a single isolated request. When something is flagged, all that gets sent back to OpenAI is a signal:
a narrowly defined safety signal, without exposing the underlying prompts or responses
Whether to dig further after getting that signal is up to the customer, who decides whether to hand over the underlying data. According to the company, the mechanism watches both people misusing the platform and AI agents that go off the rails.
One exception is spelled out plainly: images flagged as suspected child sexual abuse material are retained for human review and reporting even under zero-retention deployments, consistent with current practice.
A split with Anthropic
Another company answers the same question almost the opposite way. Multiple outlets have reported that Anthropic retains 30 days of session data for models under its coverage, with human review going through controlled access and every access logged in a tamper-resistant record. Its bet is traceability: if something goes wrong, there's a complete evidence trail, and who touched the data is easy to check.
OpenAI is betting on not producing an evidence trail at all. The fork between the two approaches isn't about safety capability — it's about what enterprise customers are actually afraid of. Companies worried about internal misuse and after-the-fact finger-pointing tend to like tamper-resistant logs; companies worried about external subpoenas and cross-border compliance would rather the data never land at the vendor in the first place — a 30-day retention window means, roughly speaking, every conversation spends a month reachable by outside legal process. For a company doing business across multiple jurisdictions, that month is a hard thing to explain to legal.
For teams buying models to go global, this can go straight onto the procurement checklist: where the data sits, who holds the keys, and what happens after a safety policy is triggered. In the past, questions like these usually got nothing more than "we take security seriously" as an answer.
Until the white paper, it's all a preview
Private Safety Processing is currently being tested with only a small set of early customers, with a gradual rollout planned for September alongside a technical white paper. The notice leaves plenty unsaid: which models are covered, exactly how fine-grained the signal is, and what appeal process a customer gets if there's a false positive. The engineering behind running cross-session analysis on encrypted storage hasn't been disclosed either — until the white paper lands, this is a promise, not a verifiable design.
For enterprise buyers, the practical move right now is to note September on the calendar. Once the white paper is out, checking it line by line against your own compliance checklist will be far more useful than parsing today's announcement.
Sources: OpenAI official announcement, CocoLoop, TechCrunch; the scope of ZDR eligibility, encryption key ownership, and the rollout plan for Private Safety Processing were verified against OpenAI's announcement.