OpenAI keeps zero data retention and previews a new safety check
On 19 August, OpenAI posted a reassurance aimed at its business API customers: Zero Data Retention stays available as its models get more capable. Zero Data Retention, or ZDR, is an arrangement for eligible customers in which OpenAI does not keep prompts or model responses after a request is processed, and its staff cannot read that content.
The reason for the post is a tension OpenAI describes openly. Some misuse only shows up when you look at many requests together, such as someone probing safeguards over and over, or spreading activity across several accounts. Safety checks that work under ZDR look at one request at a time. OpenAI also notes that some recent frontier-model deployments required customers to let the AI provider keep sensitive content for safety monitoring, which many organizations cannot accept.
How Private Safety Processing works
OpenAI's answer is a system it calls Private Safety Processing, shown as a preview. For ZDR customers, the content stays on infrastructure the customer controls. OpenAI is also building an option where content sits on OpenAI's servers, encrypted with keys that only the customer holds, so OpenAI staff cannot open it.
Automated systems then look for patterns across related requests. When they find a risk, OpenAI receives a narrow signal saying what type of activity was flagged. According to the post, its staff still do not see the prompts or responses, even after a flag. The customer can look into the alert using its own records, and can choose to share details with OpenAI to appeal a decision or to help with an investigation.
Where it stands
The system was being tested with early customers on the day of the post. OpenAI said it planned to start rolling it out in September and to publish a technical white paper then. The post does not give prices, a list of eligible plans or a date for the encrypted storage option.
What to do this week
If your company sends sensitive data through OpenAI's API, check whether your account actually has ZDR, since it applies only to eligible customers. Then put the September white paper on your calendar. The details that matter for a security review, such as what the signal contains and who can trigger an enforcement action, are not in this first post.
What we think
It is good to see a lab say in public that it wants to keep ZDR when the easier path would be asking customers to give it up. The design has a weak spot that only the white paper can address: everything depends on how much a safety signal reveals by itself. We will read that document before telling anyone the problem is solved.
Sources