Confidential Inference

RedRouter Confidential Inference protects prompts and completions while the model runs. Hardware-isolated execution, remote attestation and encryption to the verified runtime keep content private throughout inference.

How it works

Your client verifies the confidential runtime before sending a request. Prompts stay encrypted through the network edge and control plane, and are decrypted only inside the verified execution environment. The response is encrypted back to your client.

Stage	Protection
Client verification	Remote attestation checks the runtime identity and binds it to the encryption key
Encrypted request	Prompt content stays encrypted until it reaches the verified runtime
Confidential execution	The model processes content inside a hardware-isolated environment
Encrypted response	Completions return through the protected connection to your client
Operations and billing	Usage metadata supports billing and service operations without storing prompts or completions

Confidential Inference protects content from service and cloud operators outside the verified execution boundary. Verification is part of establishing that boundary; a request is not silently downgraded to ordinary inference when a confidential connection cannot be verified.

Privacy during execution and after the request

Zero Data Retention means no prompts or completions are stored. Confidential Inference adds protection while the model processes your request, using hardware isolation and an encrypted connection to the attested runtime.

Control	What it protects
HTTPS	Traffic between the endpoints of a transport connection
Encryption at rest	Stored data under its key-access policy
Zero Data Retention	Prompts and completions are not stored
Confidential Inference	Content during model execution, inside the verified hardware and software boundary

RedRouter offers both standard inference and Confidential Inference. Use the confidential integration for workloads that require protection during execution. Standard API endpoints, browser Chat and Mac app settings have their own documented behavior; contact us for the confidential client and model options that fit your workload.

Read data handling for standard API retention controls and private vs confidential inference for a fuller comparison.

What remains in your control

Your client receives the plaintext response. Local transcripts, application logs, browser extensions and any tools you invoke can create copies outside the confidential runtime. Review those destinations before sending sensitive material.

Hardware, firmware, the verified runtime and client software remain part of the trust boundary. Attestation verifies runtime identity; it does not establish that a model's answers are correct or safe. External tools and processing are covered only when they are part of the confidential integration you use.

Use Confidential Inference

Email support@redrouter.org with your workload, model requirements and integration needs. We will help you choose supported models, connect your client and confirm pricing and deployment options.

For standard OpenAI- and Anthropic-compatible requests, see the inference API. All use remains subject to our Terms of Service and privacy policy.