MorseSign inStart freeOpen Morse

What is OpenAI Private Intelligence? What it keeps, who gets it, and Private Inference

The Morse teamPublished 9 min read

As of November 2026, OpenAI Private Intelligence is OpenAI's name for a set of data protections it sells to businesses so they can send sensitive work to its models (OpenAI). It has two named parts: Zero Data Retention with Private Safety Processing, which lets automated safety reviews run "without giving OpenAI personnel access to the underlying content," and Private Inference, which "combines confidential computing with strict, verifiable controls" (OpenAI). Private Inference isn't generally available, and nothing on the Private Intelligence pages names the personal ChatGPT plans.

We read OpenAI's Private Intelligence pages, its DevDay 2026 recap, the Private Safety Processing developer guide and technical white paper, the API data controls page, and the Enterprise Key Management help article on November 18, 2026. Press coverage gives different launch dates, so we dated everything from OpenAI's own pages. OpenAI deserves credit here. Its white paper runs past 30 pages and says plainly what the design doesn't cover yet.

What does OpenAI Private Intelligence include?

Two named parts, plus related controls from the same docs. Each row is from OpenAI's own pages as of November 18, 2026, with personal ChatGPT added for contrast.

ComponentWho gets itStatus on November 18, 2026Source
Zero Data Retention with Private Safety Processing (ZDR with PSP)API organizations already approved for ZDR, set per projectRolling out to API customers "in phases" since September 22, 2026OpenAI, guide
Private InferenceCustomers "that meet the applicable eligibility and capacity requirements""Alpha" on the product page, "in preview" on the interest formOpenAI, form, white paper
Enterprise Key Management (EKM)ChatGPT Enterprise and API customersAvailable, with keys in AWS KMS, Google Cloud, Azure Key Vault, or OCI VaultOpenAI
Private Retention with PSP (formerly "Eyes Off")ZDR customers whose models OpenAI makes ineligible for ZDR, with written noticeKeeps encrypted records on OpenAI's own infrastructure; the white paper says it "is not ZDR"OpenAI, white paper
Personal ChatGPT (Free, Plus, Pro)Not named on the Private Intelligence pagesTraining data sharing on by default; temporary chats kept up to 30 daysOpenAI

The product page also groups Private Intelligence under four headings: access restrictions, data ownership, agent actions, and data security (OpenAI). Its "privacy by product" page splits the controls across ChatGPT, Codex, and the API (OpenAI).

When was OpenAI Private Intelligence announced?

At DevDay 2026, on September 29, 2026. OpenAI's news feed dates its DevDay recap that day, and the recap introduced Private Intelligence and said the Private Inference preview was "coming this fall" (OpenAI). Private Safety Processing came first. OpenAI previewed it on August 19, 2026, and added a note on September 22 that it was rolling out to API customers (OpenAI). The white paper is dated October 26, 2026 (OpenAI).

What is Private Safety Processing?

It's how OpenAI checks zero-retention traffic for misuse across many requests without its staff seeing the content. OpenAI says its earlier ZDR-compatible safety systems "evaluate each interaction individually," while some risks show up only across related interactions, such as someone repeatedly probing safeguards (OpenAI). It adds that some recent frontier-model deployments have required customers to let the provider keep sensitive content for safety monitoring, and that PSP "is designed so we can continue to offer ZDR."

The developer guide describes two steps (OpenAI):

  1. Selection and storage. An interaction is picked "through a safety classifier referral or an approved sampling policy," encrypted, and written to the customer's own cloud storage. OpenAI keeps "an index with operational metadata and a storage reference, not a copy of the content."
  2. Automated review. A "hardware-attested computing environment that disables human access" decrypts the record and reviews it. Only "predefined, bounded safety signals" leave it in plaintext.

When something is flagged, OpenAI gets "a narrowly defined signal indicating the type of activity involved," and its staff "do not receive access to the customer content even when it is flagged" (OpenAI). A customer who wants to appeal, or to help an investigation into verified abuse, can choose to share more. For how this compares with ordinary abuse logs, see what is abuse monitoring.

Does ZDR with Private Safety Processing keep your data?

OpenAI doesn't, but the customer has to. The white paper calls PSP "a distinct exception path for eligible traffic, not a synonym for ZDR" (white paper). Selected prompts and responses sit, encrypted, in a bucket the customer owns on AWS, Azure, Google Cloud, or OCI (OpenAI). OpenAI's developer guide requires customers to "retain encrypted records for at least 30 days" (OpenAI), and the white paper says OpenAI checks during setup that the bucket keeps them that long (white paper). OpenAI's ability to decrypt those records "automatically expires after 30 days," whether or not the customer deletes them.

Three more details from OpenAI's own pages:

  • Sampling can be wide. For dual-use areas such as cybersecurity, "OpenAI can sample up to 100% of traffic for asynchronous review," though the white paper says the rate varies by product and "can be significantly lower."
  • It's per project. Once on, the PSP policy "applies to all API traffic in that project, including requests to models that do not otherwise require PSP" (OpenAI). Disconnecting the storage resets the project's policy, and "models that require ZDR with PSP may become unavailable."
  • Some limits are still open. The white paper says PSP "currently does not impose a limit on the total information disclosed across repeated reviews of the same content," and that OpenAI expects to add one.

OpenAI's guide says ZDR with PSP works "without OpenAI retaining customer prompts or responses" (OpenAI). That holds for OpenAI's servers, while the customer keeps an encrypted record for at least 30 days for OpenAI's reviewer to read. For what plain ZDR keeps and what it doesn't, see what is zero data retention.

What is OpenAI Private Inference?

It's OpenAI's planned confidential-computing option, meant to protect "your prompts and responses while the model processes them" (OpenAI). Confidential computing runs code inside a trusted execution environment, which the white paper says isolates protected memory "from infrastructure outside that environment, including human access" (white paper).

OpenAI's own pages describe its status three ways. The product page labels it "Alpha," the interest form says "in preview," and the September recap said the preview was "coming this fall." We found no developer guide for it. OpenAI's index of API guides lists one for ZDR with PSP and none for Private Inference.

The white paper is narrower than the product page. It calls Private Inference "an optional protection that extends operator access protections and attestation to eligible safety-inference workloads," and says it "does not automatically cover the initial API request" (white paper). Its plans section says OpenAI expects to extend confidential computing further "when necessary technology, capacity, and product commitments are met."

Who checks the hardware proof in OpenAI's design?

OpenAI's own systems do, with auditors able to look afterward. For the PSP review runtime, the white paper says decryption keys are released only when "attestation evidence from a trusted authority satisfies the key's release policy," checked by OpenAI's key vault (white paper). Approved software versions are recorded in "a tamper-evident transparency log," and "authorized auditors can inspect" which configurations were approved. Nothing in the guide or the white paper has the customer's own software make that check before it sends a prompt. We explain who runs the check in other AI apps in how to verify an AI enclave's attestation.

What does Enterprise Key Management add to Private Safety Processing?

A second lock on the stored records. With EKM on, OpenAI's own key "is not enough to decrypt a stored record," and revoking the customer's key "prevents decryption of retained records, but does not delete them or undo completed processing" (OpenAI). OpenAI's EKM help page also says features "in beta or alpha" aren't available with EKM on, and the product page labels Private Inference "Alpha" (OpenAI). We cover how EKM itself works, and who holds the keys, in is ChatGPT encrypted.

Who can get OpenAI Private Intelligence?

Businesses, through OpenAI's API console or sales team. ZDR itself is "subject to prior approval by OpenAI and acceptance of additional requirements" (OpenAI). Only "organizations already approved for Zero Data Retention" can set up ZDR with PSP directly, and an organization administrator has to connect and validate the storage (OpenAI). The interest form asks for a work email, job title, company name, and company size (OpenAI).

Does Private Intelligence change anything for ChatGPT users?

Not for personal accounts, by anything OpenAI's pages say. The Private Intelligence pages name ChatGPT Work for enterprises, Codex, and the API, not Free, Plus, or Pro (OpenAI). OpenAI's help center says that on personal Free, Plus, and Pro workspaces, data sharing for training "is enabled by default," and that temporary chats "may be retained for up to 30 days for safety purposes" (OpenAI). Workplace plans are different, and we cover them in is ChatGPT Business private. For the personal app's settings, see is ChatGPT private.

How is Private Intelligence different from end-to-end encryption to an enclave?

The difference is who reads the prompt, and who checks the proof. Every row comes from the pages linked above, read November 18, 2026, with the Morse column from Morse's privacy policy and app.

OpenAI ZDR with PSPOpenAI Private InferenceMorse E2EE
Who reads your prompt to answer itOpenAI's serversMeant to run in confidential computing; scope not published beyond the white paperOnly the verified enclave
Who checks the hardware proofOpenAI's key vault, for its review runtime; auditors can inspect the logNot publishedYour browser, before anything leaves your device
Copy kept for safety reviewSelected interactions, encrypted, in the customer's bucket for at least 30 daysNot publishedMorse's servers relay text they can't read and don't keep messages
Who can get itAPI organizations approved for ZDRCustomers who meet eligibility and capacity requirementsPro and Max, from $19 a month

How Morse gives you the enclave check without a sales contract

Every part of Private Intelligence is a setting on a business account, and the one hardware check OpenAI has described is run by OpenAI's own systems. OpenAI's pages offer none of it to a personal ChatGPT plan. Morse is built for that person, with a privacy mode you pick for each chat.

  • E2EE, for confidential computing that your own browser checks. Your browser verifies the enclave's cryptographic proof itself, comparing what's running against versions pinned in the app, before anything leaves your device. The message is encrypted for that enclave only, Morse's servers relay text they can't read, and the reply is decrypted only in your browser. Each reply has a lock you can open to see every check that passed, with a full report to download. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models such as GLM, Kimi, and DeepSeek, with no web search or image generation, and comes with Pro at $19 a month.
  • Private, for zero data retention with no approval process. It's the default for a new chat. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
  • Anonymous, for GPT, Claude, and Gemini without your account on the request. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account, so the provider answers without knowing it's you. The provider may keep it for a limited time to check for abuse, under its own policy, so leave out names and details that point back to you.

In every mode, your chat history is stored in your browser, encrypted, not on Morse's servers. Morse doesn't train AI on your chats, sell your data, or show ads, and because it doesn't store chats, it usually has very little to give authorities (privacy policy). Clearing your browser deletes those chats, and Morse can't recover them, so keep anything you need.

Try a chat from the Morse home page without an account, or start free and open your first E2EE chat when you move to Pro.

Common questions

Is OpenAI Private Intelligence part of ChatGPT Plus?

OpenAI's Private Intelligence pages don't name Plus or any personal plan. They cover ChatGPT Work for enterprises, Codex, and the API, and the interest form asks for a work email and a company name. On Free, Plus, and Pro, OpenAI's help center says data sharing for training is on by default.

Is Private Safety Processing the same as zero data retention?

No. OpenAI's white paper calls Private Safety Processing 'a distinct exception path for eligible traffic, not a synonym for ZDR.' With ZDR with PSP, OpenAI keeps no copy, but selected prompts and responses are kept encrypted in the customer's own cloud storage for at least 30 days.

Can OpenAI staff read content that Private Safety Processing flags?

OpenAI says they can't. Its staff receive a narrowly defined signal naming the type of activity, not the content, and the customer decides whether to share more to appeal or to help an investigation.

When did OpenAI announce Private Intelligence?

At DevDay 2026. OpenAI's news feed dates its DevDay recap, which introduced Private Intelligence, September 29, 2026. Private Safety Processing was previewed earlier, on August 19, 2026, and OpenAI's technical white paper is dated October 26, 2026.

Can you verify OpenAI's Private Inference yourself?

Not with anything OpenAI has published so far. As of November 18, 2026, OpenAI's API guide index lists no Private Inference guide, and for its safety runtime the white paper says the attestation is checked by OpenAI's own key vault, with a log that authorized auditors can inspect.

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next