Secret AI
A confidential GPU worker network with a separate open attestation verifier; ordinary inference SDK calls do not integrate it.
INPUT / OUTPUT PRIVACY
Stage 1Partial proof
ASSESSED DEPLOYMENT
Direct confidential worker inference
Verification labels describe capabilities established by the cited sources. This review did not perform live cryptographic attestation, reproduce production builds or complete an independent security audit.
Who actually runs the models?
Secret AI workers run the inference engines themselves. WorkerManager registers workers and returns service endpoints; users connect to a worker rather than selecting unrelated upstream API providers. The SDK discovers worker URLs and invokes their locally served models.
Deployment and TEE coverage
Covers direct Secret AI model-worker endpoints on GPU-enabled SecretVM deployments. Applications, fine-tuning, training and the separately described private RAG store require their own assessments. Worker discovery and an API key do not establish a verified inference channel.
The worker architecture combines Intel TDX or AMD confidential VMs with NVIDIA confidential GPUs. SecretVM exposes separate CPU, runtime and GPU evidence. A CPU-only SecretVM or an unchecked worker endpoint is not proof of confidential GPU inference.
Why Stage 1?
Stage requirementsAt least one content layer has independently checkable technical evidence, including partial proof within a layer. The full privacy boundary is incomplete. The stage tooltip distinguishes partial findings from fully met requirements for Stage 2.
STEPS FOR STAGE 2
1 of 3 requirements for Stage 2 met. Missing proof: Request path, Logs & storage.
- Request pathPartial
Private inputs and outputs across the request path, with E2EE or an equivalent attested channel that protects against intermediaries and binds keys to the accepted workload.
Partial technical proof. SDK calls remain unchecked
- Inference executionMet
Protected inference with a verifiable workload identity and protection covering the CPU, GPU and every place content is processed.
Technically verifiable. CPU + GPU; workload replay
- Logs & storageNot established
No content logging, persistent storage, human review or training. Any temporary content-derived cache must have verifiable isolation and enforced removal.
Documented only. Worker lifecycle unverified
Identity and operational metadata are assessed separately below. They do not set the content stage; prompt or response content in telemetry remains part of the content-retention assessment.
Request path
Partial technical proofSeparate verifier available. The separate open-source secretvm-verify client checks CPU/GPU evidence and TLS-key bindings, and accepts a distinct service TLS URL so the inference-port certificate can be checked rather than only the attestation server. This corrects the earlier implication that no verifier exists. ChatSecret.get_attestation() still returns an empty object; ordinary inference calls do not integrate those checks or pin the subsequent connection. Complete serving-port binding, accepted workload policy and fail-closed inference integration remain deployment/client responsibilities.
Inference execution
Technically verifiableWorker evidence available. SecretVM publishes reproducible VM measurement procedures. Its separate verifier validates TDX/SEV-SNP and NVIDIA evidence, checks CPU/GPU nonce bindings and replays the workload Compose measurement against a signed build registry. Require GPU evidence for a GPU worker and review every accepted image/configuration; a CPU-only success or authentic Compose containing an unpinned image is insufficient. This establishes available verification capability, not proof that every discovered worker has passed those checks.
Logs & storage
Documented onlyZero-logging claim. The current Confidential AI product page promises zero prompt logging for its preconfigured models and advertises cached-input pricing. That is a documented claim, not proof that no content-derived cache survives a request. The reviewed worker architecture and SDK do not establish enforced log suppression, tenant cache isolation/removal, backup behavior or training exclusion for every endpoint. Arbitrary SecretVM applications and the separate confidential PRAG store require their own lifecycle assessments.
Identity & metadata
Documented onlyAccounts and protected usage. Workers report token counts for subscription validation and settlement. Secret AI describes model access, payments and earnings as protected Secret Network contract data. Separately, the current portal policy collects account details, IP/device information, SDK/deployment logs and session metadata, with third-party payment processing and purpose-based retention. Protected contract records do not establish unlinkability from worker connections or deletion of off-chain account and usage records. These metadata practices do not set the content stage.
Remaining trust assumptions
- Intel or AMD VM isolation and NVIDIA GPU isolation, with both evidence chains checked for the selected worker.
- WorkerManager discovery, the chosen worker's measured application configuration and its update policy.
- The user's verifier and binding of the inference connection to the approved worker, beyond ordinary SDK authentication.
Limits of this assessment
- The inference SDK attestation accessor remains a stub; the separate SecretVM verifier must be integrated with the actual serving connection.
- A valid CPU quote or signed Compose file alone does not establish accepted GPU state, immutable model-serving images or content deletion.
- No live worker quote, serving connection or independent deployment audit was verified in this review.
Primary-source record
9 sources · reviewed 2026-10-09
Sources reviewed 2026-10-09