On August 19, 2026, OpenAI confirmed it is extending zero data retention (ZDR) to its most advanced models, alongside a new abuse-detection system called Private Safety Processing. For any SMB feeding customer data, contracts, or internal documents into ChatGPT or the OpenAI API, this announcement touches a sensitive question directly: where does that data go, and who can read it?
In brief
- On August 19, 2026, OpenAI announced it will continue and expand zero data retention for its frontier models, limited to API customers deemed "eligible."
- A new system, Private Safety Processing, is meant to detect abuse patterns across multiple interactions without OpenAI staff reading the content of conversations.
- Broader rollout and a technical white paper are planned for September 2026; eligibility criteria have not yet been specified.
- Acknowledged limitation: retention is never fully zero — content flagged as related to child-abuse material is still kept to meet legal reporting obligations.
- According to Gartner, data governance remains the top barrier companies cite when adopting generative AI.
What OpenAI announced on August 19, 2026
OpenAI confirmed it will keep offering zero data retention for its frontier models to "eligible" API customers. In practice, the company states it does not store prompts and responses once a request has been processed, does not make that content accessible to staff for review, and does not use it to train its models unless the customer explicitly opts in.
The announcement comes with a second piece: Private Safety Processing (PSP), a system now being tested with pilot customers. Its goal is to catch abusive usage patterns that only become visible across several linked interactions, without giving OpenAI staff access to the raw content of those exchanges. OpenAI frames this as a response to how the models are now used: the more autonomously and the longer AI agents work, the more some risks only surface when interactions are cross-referenced.
In its communication, OpenAI summed up the stakes this way: enterprise customers see data privacy as non-negotiable.
Request sent
Processed, not stored
Abuse detection without human review
Legal reporting if required
Why this announcement, now
The timing makes more sense in context. Gartner identifies data governance as the top barrier to enterprise AI adoption, ahead of cost or model reliability. IT departments hesitate to connect sensitive data (contracts, HR files, customer records) to tools whose handling of that data they don't fully control.
The announcement also lands amid growing competition among AI vendors over the privacy guarantees they offer businesses. Several analysts quoted in trade press also read it as a move that positions OpenAI favorably ahead of a possible IPO, reportedly discussed internally for 2027 according to remarks attributed to CFO Sarah Friar. That reading remains an analyst interpretation, not a fact confirmed by OpenAI itself.
Before August 2026
Zero retention already existed
August 19, 2026
Public announcement
September 2026
Broader rollout planned
What "zero" covers, and what it doesn't
The term "zero retention" is worth unpacking to avoid confusion. The table below summarizes what standard ChatGPT/API usage covers compared with a zero data retention arrangement.
| Criterion | Standard usage (API/ChatGPT) | Zero data retention (ZDR) |
|---|---|---|
| Prompt and response storage | Yes, for a duration set by OpenAI's policy | No, deleted after processing |
| Use for training | Possible unless disabled by the customer | No, unless the customer explicitly agrees otherwise |
| OpenAI staff access to content | Possible for standard moderation | No, except mandatory legal reporting |
| Eligibility | Open to most paying customers | Limited to "eligible" API customers, criteria unspecified |
| Absolute exception | Legal reporting in cases of serious abuse | Same: legal reporting in cases of serious abuse |
Worth remembering
"Zero" is never absolute. Even under a ZDR contract, content flagged as related to child-abuse material is still retained to meet legal reporting obligations. And as of today, OpenAI has not published the exact criteria that make a customer "eligible" for zero retention — that answer is promised for September 2026.
What this concretely means for an SMB
For an SMB, this announcement doesn't open automatic access to zero retention. The arrangement is, for now, limited to API customers deemed eligible, a notion OpenAI hasn't publicly defined yet. Three practical steps help while waiting for more clarity.
First, check your current contract with OpenAI (or any AI vendor) rather than assuming a protection that doesn't exist by default. Standard retention applies until a specific option has been activated or negotiated.
Second, don't wait until September 2026 to act on genuinely sensitive data. An SMB handling customer files, health data, or trade-secret information should already limit what it feeds into a general-purpose AI tool, ZDR or not, and favor internal channels validated by its AI policy.
Third, track the details expected this fall: eligibility criteria, possible pricing for the ZDR option, and technical specifics of Private Safety Processing. An SMB negotiating an enterprise contract with OpenAI, or a competitor, can reasonably ask for these guarantees in writing rather than relying on a general announcement.
Without zero retention
Prompts and responses may be retained under the vendor's standard policy, and potentially reused for training unless explicitly disabled.
With zero retention (eligible customer)
Content is not stored after processing nor used for training, unless otherwise agreed. Only mandatory legal reporting is an exception.
With measured optimism, this move points in the right direction: AI vendors are responding to a real business demand around privacy, rather than treating it as an afterthought. The open question, come this fall, is whether the eligibility criteria make this promise accessible to SMBs, or only to the largest accounts.
FAQ
What is zero data retention?
It's a processing mode in which an AI vendor does not store prompts and responses once a request is processed, does not make them accessible to staff for review, and does not use them to train its models, unless the customer agrees otherwise.
Can an SMB request zero retention from OpenAI right now?
The option already exists for some API customers, but OpenAI limits zero retention to customers deemed "eligible" without having published the exact criteria. Details are expected in September 2026, alongside a technical white paper.
Does Private Safety Processing still read conversation content?
OpenAI states this system detects abusive usage patterns across several linked interactions without exposing the raw content to a human operator. Precise technical details on how it works have not yet been published.
Is zero retention really total?
No. Even under ZDR, content flagged as related to serious child-abuse material is still retained, to meet a legal reporting obligation. That is the only exception OpenAI acknowledges.
To learn more about AI security and compliance for businesses, see our guide AI Vendor Security: The 2026 Ranking, our article on Shadow AI in SMBs, and all our resources on AI for business.
