Inputs and permissions
How encrypted inputs and key access work, how each required owner authorizes independently, and how each operation is approved when it is requested.
Agents need authority for the resources they use and the results they release. Standard binds those permissions to execution, whether the resources belong to one organization or several owners.
Each required owner authorizes independently. One owner’s approval cannot replace another’s, and a sponsor or coordinator cannot approve on an owner’s behalf. Owners sign their approvals and release authorizations with keys they hold.
In Conversation analytics, the conversation batch, classifier and reporting policy have distinct roles. The Analytics agent can request counts without gaining authority to change the classifier, expand the batch or choose another recipient.
Choose an input route
For local files, the participant host encrypts inputs before contribution to the accepted confidential runtime. For connected sources, the request identifies an approved connection and resource selection; attested confidential ingress retrieves and validates the material before sending it to the operation worker. Private contents stay out of agent conversations and ordinary tool arguments. See Sources and customer vaults for the complete flow.
An encrypted-input reference identifies an object and its expected digest. Holding the reference does not grant permission to decrypt or use it. References can still contain sensitive metadata and should follow your disclosure policy.
Key access and attestation
The key-access policy determines which attested runtime and authorized execution may obtain decryption material. Required checks include the accepted runtime identity, workload, and applicable authority. If the required conditions are not met, the protected execution cannot use the input.
Non-exportable vault/KMS master keys remain in the customer’s key store, where decrypt or sign operations use them. Transient data keys and permitted API secrets can be used in confidential memory. Customer-store policies govern access; provider-specific authentication and attestation conditions apply. Agent requests carry references rather than raw credentials.
Code, models, and validation inputs
Code and model artifacts can be protected inputs as well as data. They must match the runtime's supported formats and resource limits.
Accept the classifier and its inputs under authority separate from the requesting agent. Hardware attestation alone does not establish classification accuracy. The accepted policy identifies the required function, batch and output; evidence binds the result to that operation. Independent checks of agent-written code are covered in the separate release-verification example.
See Formats and limits.