Phala Cloud Documentation — Confidential AI on TEE

Phala Confidential AI is designed so privacy and integrity claims can be checked with technical evidence: attestation reports, source provenance, signed receipts, and provider verification records.

Read Trust Boundary first. The gateway is attested and receipts do not retain request bodies, but plaintext is visible inside the attested gateway after TLS termination or E2EE (end-to-end encryption) decryption.

Technical Controls

Control	How Confidential AI provides it
Data isolation	Gateway and confidential providers run in TEEs.
Hardware-backed identity	The gateway publishes an attestation report with a TEE quote.
Per-response auditability	Receipts bind request and response hashes to the attested workload.
No body retention in receipts	Receipts store hashes and facts, not prompt or response text.
Upstream zero data retention	`provider: {"zdr": true}` restricts routing to providers that do not retain prompt or completion content. Enforced fail-closed, including on failover. See Zero Data Retention.
Confidential upstreams	For confidential responses, the upstream is verified and channel-bound before forwarding.
Encryption in transit	TLS for all requests, with optional E2EE.

Regulatory Fit

Requirement	How the architecture helps
Data minimization	Gateway receipts avoid storing request and response bodies. Request `zdr` to extend the constraint to the upstream provider.
Storage limitation	Zero-data-retention providers do not keep prompt or completion content after serving a request.
Confidential processing	TEE isolation and attestation provide verifiable runtime identity.
Audit evidence	Attestation reports, receipts, sessions, and source provenance can be retained by the customer.
Provider separation	Receipts distinguish confidential upstreams from routed third-party providers.

For regulated data, use confidential responses and verify upstream.verified.result = verified, required = true, and the provider claims your policy requires.

Source Provenance

Every gateway attestation report includes attestation.source_provenance:

{
  "repo_url": "https://github.com/Dstack-TEE/private-ai-gateway.git",
  "repo_commit": "9d45c7e3d48d2f74c31cd85f1fb5c6cee1435ef3",
  "image_digest": null,
  "image_provenance": null
}

Use these fields to compare the running workload against the source release you are willing to trust.

Relevant Repositories

private-ai-gateway

ACI gateway, attestation reports, receipts, and provider verification.

dstack

TEE runtime and KMS used for workload identity and quote generation.

What to Audit

Attestation report generation and nonce binding.
Keyset endorsement and receipt signing keys.
Provider verification and channel binding.
Receipt event log construction.
Fail-closed behavior when confidential upstream verification is required.
Source provenance and image provenance for production releases.

Enterprise Documentation

Compliance reports, enterprise security reviews, DPAs, BAAs, and certification scope can change over time. Request current documents from Phala through your account or enterprise support channel.

Confidential AI

Compliance and Open Source

Technical Controls

Regulatory Fit

Source Provenance

Relevant Repositories

private-ai-gateway

dstack

What to Audit

Enterprise Documentation

Attestation Report

Receipts and Sessions

​Technical Controls

​Regulatory Fit

​Source Provenance

​Relevant Repositories

private-ai-gateway

dstack

​What to Audit

​Enterprise Documentation

​Related

Attestation Report

Receipts and Sessions

Technical Controls

Regulatory Fit

Source Provenance

Relevant Repositories

What to Audit

Enterprise Documentation

Related