Corporate confidentiality put to the test by advanced models

At DevDay 2026 in San Francisco, OpenAI unveiled OpenAI Private Intelligence, an offering aimed at companies that handle sensitive information. It pairs zero retention of input and output data with Private Safety Processing, an automated monitoring system designed to spot abuse spread across multiple conversations. The first component, ZDR with Private Safety Processing, is already rolling out; the second, Private Inference, meaning inference inside hardware enclaves with verifiable controls, is for now only a preview announced for this autumn, with no date and no access requirements. Back in August, when it introduced Private Safety Processing, OpenAI justified these controls by noting that the most serious safety risks are not always visible in a single interaction.

Integration comes with specific requirements. The official documentation asks customers to set up storage under their own control and to meet further technical and operational requirements. There are two deployment options: the customer's own infrastructure, or OpenAI servers with content encrypted using keys held by the customer. The same documentation sets out an exception for images: if the classifier detects possible child sexual abuse material, the image is retained for manual review even with ZDR enabled. The approach differs from the competition's: where OpenAI aims to keep ZDR by entrusting oversight to automated systems and customer-held storage, Anthropic has, since 9 June 2026, retained prompts and responses from its flagship models for 30 days, including those used by customers with immediate-deletion guarantees.

The technical picture, however, still has pieces awaiting confirmation. Some details, such as the rollout starting on 22 September, the 30-day retention of encrypted records in the customer's storage and the choice between AWS, Azure and Google Cloud, come from press reports, and we were unable to confirm them in official OpenAI sources. The announced technical white paper does not appear in the available sources, and there is no indication that independent bodies have so far audited the guarantees claimed for confidential inference.

Trying to protect confidentiality by handing oversight to automated systems and customer-held storage is a complex bet. How sustainable this balance is beyond what the documentation claims remains to be seen. — Olya