A security team gives an AI assistant an incident report. The report names affected systems, suspected weaknesses and the people investigating them. The team wants a useful answer. It also wants the service to detect abuse without turning that investigation into another source of readable intelligence.

That is the problem behind OpenAI’s Private Safety Processing, or PSP. Safety analysis sometimes needs to connect several interactions. Preventing staff from browsing those conversations is valuable. But software can still extract sensitive facts, even when no person sees the original report.

Mirror protects the computation and the release together. Our encrypted inference keeps supported analysis on ciphertext. Runtime policy binds its use to the approved purpose, and release controls govern which results can leave. The incident team receives a safety signal without giving the remote FHE evaluator a readable report.

The conversation gets a second processing path.

PSP is separate from the model answering the original request. In the documented ZDR configuration, selected interactions become encrypted records in customer-controlled storage. An approved, hardware-attested runtime retrieves and decrypts them for safety review. It releases bounded signals; detailed results return to encrypted storage. OpenAI’s implementation guide.

PSP: protect the reader and restrict its output.The safety-review path, after the original model interaction.
  1. Selected interactionPrompt and response become a safety record.
  2. Customer storageEncrypted record waits for authorized review.
  3. Approved runtimeDecrypt. Read. Analyze.Plaintext exists here.
  4. Bounded signalOnly permitted fields leave in plaintext.

The whitepaper describes substantial controls: attested key release, restrictions on network access, runtime-enforced output rules, and reviewer prompts that require a second approver and an entry in a transparency log. This is more than a promise that employees will not look. PSP whitepaper, PDF pages 20–21 (printed 17–18).

The same paper lists extending confidential-computing protection to the remaining plaintext-handling components as future work. The protected inference step must therefore be assessed alongside the rest of the safety pipeline. Planned enhancements, PDF page 33 (printed 30).

It still leaves two questions to answer separately: what can the approved reader see, and how much can its approved answers reveal?

A small answer can reveal a private fact.

Suppose an internal safety dashboard returns only a count. It never returns a transcript, a name or a sentence from the incident report. Ask about four reports, then ask about the same group with one report removed. Subtract the answers. The result can reveal the excluded report’s classification.

Two permitted-looking answers. One extra disclosure.

Exact-count illustration. This is not PSP’s output schema or a test against OpenAI.

First queryReports A, B, C, D2 flagged
Next queryReports A, B, CNot requested

The first count alone does not identify whether report D was flagged.

This is the difference between limiting an answer’s format and limiting what someone can learn. Encryption at rest does not solve it. Confidential computing does not solve it by itself. FHE protects the computation; release policy governs what its answer reveals. An encrypted computation can still produce an overly revealing answer.

The PSP whitepaper explicitly identifies the gap. It says the system does not yet bound total disclosure across repeated reviews of the same content and lists cumulative protection as future work. That is a documented design limit, not a speculative breach allegation. “Limits Across Repeated Reviews,” PDF page 33 (printed 30).

Mirror enforces release policy before a result leaves the workflow. The policy binds the approved source, operation and output. In the workflow below, it rejects an altered cohort and reuses the approved release noise on retries. Signed evidence records the release that was permitted.

Key ownership and permission to use data are separate controls.

Customer key ownership matters. So does the authority granted to use that key. OpenAI’s EKM keeps the master key in the customer’s KMS, while authorized services can obtain working keys for stored data. Those working keys can be cached in memory; the FAQ says revocation typically takes effect within an hour. EKM key lifecycle.

PSP adds its own inner encryption and approved-runtime key-release checks. Removing the customer-controlled outer layer alone does not expose the safety record. The question is which combination of permissions, software and keys can make it readable. PSP encryption layers.

Mirror separates the evaluator from the decryption key. On the FHE path, the evaluator receives ciphertext; decryption stays at the configured client or service endpoint. Our runtime controls separately check whether the requested use and destination are authorized. Possession of an encrypted record does not grant permission to read or repurpose its result.

The gaps Mirror addresses.

Mirror brings encrypted inference and runtime enforcement into the same workflow. FHE protects the supported computation; AgentIQ applies policy to the requested action; release controls govern the result; signed receipts bind the released artifact to its authorization. These controls address different ways that a private safety review can expose information or exceed its purpose.

Mirror computes on ciphertext and enforces release policy.From an approved analysis to a verified release.
  1. Authorize the analysisPolicy binds the source, purpose and permitted result.
  2. Evaluate ciphertextThe supported FHE computation runs on encrypted inputs.No decryption key at this evaluator.
  3. Control disclosureRelease policy keeps case details inside the authorized boundary.
  4. Verify the resultA signed receipt binds the released artifact to its authorization.

For the incident team, the safety check produces a useful signal while the report remains encrypted at the remote evaluator. Our controls check authorization when the source records, intended use or destination change.

The privacy gapWhat Mirror provides
Approved analysis still exposes readable content to its evaluator.FHE inference computes on ciphertext without giving the remote evaluator the decryption key.
Permission for one review becomes permission for another use.Runtime authorization binds the operation to the approved source, purpose and scope. A permitted review does not automatically authorize a new derivative or action.
Repeated results reveal an individual record.Release policy controls the approved cohort and output. Our published workflow rejects a changed corpus and reuses its release noise on retries, preventing those retries from averaging away the noise.
An agent sends a valid result to an unapproved destination.Destination-side checks enforce the approved recipient and current authority before delivery.
The recipient cannot verify which result was authorized.Signed receipts bind the released aggregate and derivative to the approved manifest, so the recipient can verify their lineage without receiving the source records.

Run Mirror’s protected safety workflow.

Our published private-oversight workflow uses four fictional records. It classifies them through encrypted model calls, keeps the individual labels inside the harness, and releases an approved noisy aggregate. A second protected step uses that aggregate to select a fictional evaluation case. The browser verifies the signed lineage.

Then it tries a changed corpus. The request is rejected instead of generating another answer over a smaller group. Repeating the same approved manifest reuses the release noise, preventing that repeat from averaging away the noise.

Mirror protects the analysis and its results.

PSP protects safety review through approved software, attested key release and restricted outputs. Mirror adds a different protection: supported analysis runs without giving the remote evaluator plaintext, while runtime policy controls the use and release of its result.

Mirror provides encrypted computation, purpose-bound authorization and controlled release today. These capabilities work together through ETF and Mirror’s enterprise platform. The evaluator performs the approved computation without its decryption key. Mirror’s surrounding controls govern which result becomes readable, where it may go and what evidence accompanies its release.

For the separate retention question, read our ZDR analysis. For state that survives a request, see Beyond the prompt. This article’s question is what happens while private data is being used, and what remains knowable afterward.

Sources and evidence

PDF references checked .