OpenAI Private Intelligence: ZDR, PSP & GDPR for Enterprises

👁 120 reads

OpenAI Private Intelligence is a new enterprise offering announced at DevDay 2026: Zero Data Retention combined with automated security checks without human involvement (Private Safety Processing), and this fall, Private Inference - protection of data during processing itself. We explain how it works technically, who can use it, and what it might mean for your company in the context of GDPR and the EU AI Act. The legal section is our analysis, not legal advice.

What OpenAI Announced

In the DevDay 2026 recap (September 29), OpenAI describes Private Intelligence as a way for companies to use the most powerful models with greater confidence that their data is protected. Sam Altman said on stage, according to CNBC’s coverage, that ZDR with Private Safety Processing provides model security “without storing content on our servers,” and Private Inference will ensure privacy at the moment of processing. OpenAI worked on the project with partners including Cisco, Databricks and Snowflake. There’s no pricing or open enrollment yet: companies express interest through a form. We cover the rest of the evening’s announcements in our live coverage of OpenAI DevDay 2026.

How ZDR with Private Safety Processing Works

By default, OpenAI’s API stores abuse monitoring logs for up to 30 days. Zero Data Retention removes customer content from these logs, but requires prior approval from OpenAI. Private Safety Processing (PSP) aims to reconcile two things: abuse monitoring and not storing content at OpenAI. The documentation bases PSP on three principles:

1. Customer controls content. Records go to the customer’s storage (S3, Azure Blob, Google Cloud Storage or OCI Object Storage), and OpenAI keeps only an index with metadata and a reference.
2. No human review. Content is decrypted only by the Safety Review Runtime, a hardware-attested environment without human access. Only limited safety signals and metadata leave the environment.
3. Storage only for safety. Content cannot be used to train models or shared with other OpenAI teams or partners.

Not every conversation is recorded. A record is created when flagged by a safety classifier or an approved sampling policy, and according to OpenAI, flagging alone doesn’t mean a policy violation. Records use OpenAI-managed HPKE encryption for the review environment. OpenAI recommends enabling Enterprise Key Management (EKM) to add a customer-managed encryption layer. When EKM is enabled, revoking its authorization blocks decryption, but doesn’t delete records or reverse completed analysis.

Who Can Use This

Two important limitations. First, Agents API according to the documentation doesn’t support ZDR and keeps data only in the US, so agents with computer use aren’t covered by this protection. Second, OpenAI reserves the right to disable ZDR for specific customers for selected models (with written notice). In that case, content is stored encrypted at OpenAI without human review, unless required by law. In a separate “Safety Retention” mode (also with notice), for serious risks OpenAI may retain content and review it manually.

Private Inference is meant to combine confidential computing with “strict, verifiable controls.” As of September 29, OpenAI hasn’t published technical documentation, so we don’t know what hardware it will run on or how verification will work.

GDPR: What This Changes for Your Company

This is not legal advice. Below is our analysis based on OpenAI’s documentation. Before implementation, consult with a data protection officer or lawyer.

• Minimization and retention (Art. 5 GDPR). PSP genuinely reduces the amount of content stored at OpenAI. However, encrypted records sit in your storage for 30 days, so they should appear in your record of processing activities and retention policy.
• Key control. With EKM, access to records depends on your key. This is an argument in impact assessments (DPIA), but the documentation explicitly states that revoking the key doesn’t delete data.
• Transfer outside EEA. PSP doesn’t move processing itself to Europe. That’s what EU data residency is for (eu.api.openai.com, EEA and Switzerland), which according to the documentation requires ZDR, modified abuse monitoring or one of the PSP modes, and costs 10% more for models released from March 5, 2026. GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna work with EU residency in Standard, Flex and Batch modes. Fast mode doesn’t work with EU residency, and Ultrafast only works in the US.
• OpenAI’s role. OpenAI still processes data when generating responses. The data processing agreement (DPA) and legal basis for processing remain your responsibility.

For comparison: the same day we wrote that Microsoft subcontractors manually review images from Copilot. PSP is designed to exclude such human review of customer content, but only in the API and only for companies with ZDR.

AI Act: What Doesn’t Change

Private Intelligence concerns data protection, not risk classification. AI Act obligations depend on what you use the model for, not where the logs sit. From our AI Act guide: obligations for general-purpose model providers rest with OpenAI, and the company using the model is the deployer. Transparency requirements (informing people they’re talking to AI) apply from August 2, 2026. National authorities can audit and fine from October 28, 2026. Requirements for high-risk systems, such as AI in recruitment, take effect December 2, 2027.

Our opinion: for companies in regulated industries, PSP and EKM will make conversations with security teams and auditors easier, because they show who holds the key. However, they won’t replace DPIA or documentation required by the AI Act.

Frequently Asked Questions

How much does OpenAI Private Intelligence cost?

OpenAI hasn’t published pricing. Companies express interest through a form on openai.com. EU data residency costs 10% more for newer models.

Does Private Safety Processing work in ChatGPT Enterprise?

The documentation describes PSP for the API and projects in the OpenAI console. We found no information about ChatGPT Enterprise.

Is ZDR with PSP enough to comply with GDPR?

Not by itself. It reduces the amount of data at OpenAI, but controller obligations, the data processing agreement and any transfer outside the EEA remain the company’s responsibility. This is not legal advice.

When will Private Inference be available?

OpenAI announced a preview “this fall.” As of September 29, 2026, there was no date or documentation.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

PSP storage and customer decryption

OpenAI’s PSP guide lists Oracle Cloud Infrastructure Object Storage alongside AWS S3, Azure Blob and Google Cloud Storage as options for customer-controlled storage. The storage region must meet the project’s data residency requirements.

The guide also describes a separate PSP decryption feature for accessing customer content in stored records. OpenAI uses a customer-controlled public key, while the private key remains in the customer’s cloud key-management service. This feature is separate from EKM, which adds a customer-managed encryption layer. Registering a decryption key does not enable PSP or configure storage. OpenAI says to confirm availability and activation with an OpenAI contact; existing records are not converted for the key.

Can a customer decrypt PSP records?

Yes, for records written with PSP decryption enabled for the customer’s key, subject to availability and activation.

Does registering a decryption key activate PSP?

No. The storage connection and PSP configuration must be set up separately.

Sources for this update

• https://developers.openai.com/api/docs/guides/private-safety-processing

Kinetyra