OpenAI formally commits to zero data retention for API customers and previews private safety processing
OpenAI reaffirms Zero Data Retention for eligible API customers and previews Private Safety Processing, a cross-interaction safety system launching in September 2026.
On 19 August 2026, OpenAI published a formal commitment to Zero Data Retention (ZDR) for eligible API customers, alongside a preview of a new system called Private Safety Processing. Together, these two things address a tension that has been building in enterprise AI procurement: how do you let a model provider run safety checks on your API traffic without handing over your users’ sensitive data?
What OpenAI is actually promising
Zero Data Retention is not new. OpenAI has offered it to certain API customers since 2023. What changed on 19 August is the explicitness of the commitment and how it fits into OpenAI’s broader enterprise offer.
Under ZDR, OpenAI does not retain your prompts or model responses after a request is processed. Your content is not available to OpenAI personnel for review, and it is not used to train their models unless you explicitly opt in. For deployments using ZDR, customer content stays on infrastructure the customer controls. OpenAI is also developing an option where content is stored on OpenAI infrastructure but encrypted with keys held by the customer.
On the API mechanics: under ZDR, the store parameter for /v1/responses and /v1/chat/completions is always treated as false, regardless of what your request sets. Image generation is ZDR-compatible across gpt-image-2, gpt-image-1.5, gpt-image-1, and gpt-image-1-mini.
One legal carve-out applies regardless of your ZDR status. Images flagged for potential child sexual abuse material (CSAM) are retained for manual review and legal reporting, as required by law. This applies to all frontier model providers and is not specific to OpenAI.
The problem ZDR alone cannot solve
Safety monitoring has traditionally worked by logging interactions and reviewing them, either automatically or with human oversight. That model breaks down under ZDR because there is nothing to review.
For most types of abuse, looking at a single prompt-and-response pair is enough to spot a problem. But Aleah Houze, Head of Product Policy at OpenAI, explained the emerging challenge plainly: “We’re seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions.”
A bad actor trying to engineer malware, for example, might spread requests across multiple sessions specifically to avoid triggering single-interaction detection. Private Safety Processing is designed to close that gap without requiring OpenAI to see what was actually said.
How private safety processing works
OpenAI is deliberately sparse on technical details ahead of the September white paper, but the broad mechanism is clear. Private Safety Processing analyses patterns across related interactions without exposing the underlying prompts or responses to OpenAI personnel.
If the system detects something concerning, it sends OpenAI a “narrowly defined safety signal” that identifies the type of potential activity, not the content itself. OpenAI can then decide whether enforcement action is needed. If it is, OpenAI contacts the customer for context, and the customer chooses whether to share additional data. Nothing is surfaced to OpenAI automatically beyond that signal.
Industry observers are speculating that the underlying techniques likely involve ephemeral processing, homomorphic encryption, or on-device safety models. OpenAI has not confirmed any specifics. The technical white paper, expected in September 2026, should fill in those gaps.
What this means for you
If you are an enterprise or API customer in a regulated sector, particularly healthcare or finance, this matters in a fairly direct way. Some recent frontier-model deployments have required organisations to allow their AI provider to retain sensitive content for safety monitoring. For many, that conflicts with data protection obligations or commitments to the people they serve. ZDR with Private Safety Processing is OpenAI’s answer to that conflict.
If you are on a consumer plan, Free, Plus, Go, or Pro, none of this applies to you. OpenAI’s ZDR controls are an API-tier offering, and your existing consumer data settings remain unchanged.
If you are evaluating OpenAI against Anthropic for enterprise use, the competitive context is worth knowing. Anthropic stipulates that for its most capable “covered” models, user sessions and conversations may be retained for up to 30 days for safety analysis. OpenAI is making a different bet: that automated cross-interaction safety analysis can provide equivalent safety guarantees without requiring any data retention at all. Whether the September white paper validates that claim technically will be worth watching.
A note on the broader context
Some analysts have been direct about the timing. Carmi Levy, an independent technology analyst, described the announcements as “a slickly conceived move to win PR points as safety concerns around agentic AI continue to mount,” and noted that positioning OpenAI favourably ahead of a potential IPO is an obvious parallel motivation.
That scepticism is reasonable to hold. At the same time, the specific technical commitment, cross-interaction safety analysis without content access, is a meaningful engineering problem regardless of what else is happening around it. The September white paper will be the real test of whether the implementation matches the framing.
OpenAI has said it will keep customers informed as Private Safety Processing rolls out, sharing updates early and explaining what they mean for existing contractual commitments. That is worth holding them to.