Chutes
A distributed inference network with model-dependent confidential computing and an optional encrypted request path.
INPUT / OUTPUT PRIVACY
Stage 1Partial proof
ASSESSED DEPLOYMENT
Public LLM API; E2EE is optional
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?
Miners execute deployed model containers using Chutes software. Chutes supplies the validator, serving API and deployment system; this is a compute network rather than a directory of unrelated upstream APIs.
Deployment and TEE coverage
The assessment covers public system-user LLM inference, including the standard API path. Private/user-created chutes and chat-history storage have different guarantees.
The broader platform supports TEE and non-TEE deployments. On 2026-10-06, the public /v1/models snapshot advertised 14/14 models with confidential_compute: true. These are provider-reported flags, not verified hardware reports, and do not cover custom chutes. Check the live flag for the exact model; a -TEE name alone is insufficient.
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
0 of 3 requirements for Stage 2 met. Missing proof: Request path, Inference execution, Logs & storage.
- Request pathNot met
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.
Operator access. Standard gateway sees content
- Inference executionPartial
Protected inference with a verifiable workload identity and protection covering the CPU, GPU and every place content is processed.
Partial technical proof. CPU + GPU quotes; workload gap
- 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. Public API only
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
Operator accessE2EE optional. In standard API mode, the gateway can read request bodies transiently. The optional local proxy encrypts payloads to an instance and checks confidential_compute by default. Chutes now publishes nonce-bound evidence and an E2EE public-key binding, so a custom verifier can check the destination key. The reviewed transport discovers keys without performing that complete attestation and approved-measurement check itself. Optional encryption does not protect the standard API path assessed here.
Inference execution
Partial technical proofCustomer evidence available. Customers can retrieve fresh per-instance Intel TDX and NVIDIA evidence, check hardware signatures and debug state, and bind the E2EE key through report_data. Public sek8s tooling can reproduce VM measurements. This corrects the earlier validator-only description. The sample verifier does not enforce an independently accepted measurement policy, and this review did not establish complete model-container identity, configuration and source linkage for every eligible serving instance. Catalog flags alone remain insufficient.
Logs & storage
Documented onlyZero-content policy. Chutes states that public system-user API prompts and responses are not persisted; only short-lived prefix hashes and usage summaries support routing and billing. Public serving code also contains diagnostic log files and an authenticated log-reading service. Those facilities are not proof that prompts are logged, but the policy and presence of attestation do not establish the deployed log fields, protected cache isolation or enforced cache-removal lifecycle. User-created chutes and saved chat history are separate scopes.
Identity & metadata
Documented onlyUsage retained. Usage and billing metadata are retained. Chutes says core databases do not store IP addresses, but security infrastructure may process them. Metadata retention is purpose-based without a complete fixed schedule or verifiable deletion controls.
Remaining trust assumptions
- Intel and NVIDIA isolation and attestation roots for TEE models.
- Chutes validator, approved serving code and model configuration.
- The locally operated E2EE proxy and authenticity of the destination instance key when E2EE is used.
Limits of this assessment
- TEE coverage is model-specific and may change; the dated public catalog is narrower than the platform.
- Standard API calls expose plaintext at the gateway; optional E2EE requires checking and pinning the attested destination key and workload.
- Open repositories and valid hardware quotes do not by themselves establish the full serving implementation or content-removal policy. No live quote was verified.
Primary-source record
9 sources · reviewed 2026-10-09
Sources reviewed 2026-10-09