Back to articles

Secrets without handing them to the model

Keep secret material encrypted and use scoped expressions so machines can act without leaking credentials into templates, logs, or AI context.

Kleis encrypted secret reference flowing through policy to a workloadTasmanian Cloud kleis encrypted secret reference flowing through policy to a workload.[KLEIS]SECRET REFno plaintextPOLICYscoped accessWORKLOADdecryptsKeep the boundary explicit. Let each system do the work it is good at.

The .env file is a boundary failure

Environment variables are useful at the last step of a process. They are a poor system of record. A value copied into a .env file can move into a repository, a shell history, a support bundle, a generated Compose file, or a model transcript.

Kleis is designed to keep the value behind a reference until the authorised machine has a reason to use it.

Dialkeys name the service relationship

A dialkey gives a service or deployment a logical identity that can be used for discovery and derived configuration. The workflow can issue one identity, derive purpose-specific values, and publish the service without asking a customer to hand-maintain a pile of unrelated environment variables.

Dialkeys are reachability and service identity inputs. They do not, by themselves, grant a tenant access to another tenant.

Kleis evaluates the use

Kleis expressions let a workflow refer to a credential, token, or key without copying the plaintext through every layer. Access can be scoped to a workload, environment, action, or time window.

The control plane can carry and audit the reference. Decryption or use happens at the boundary where an authorised machine needs it.

Machine-to-machine without shared secrets

A service can request a narrowly scoped operation instead of receiving a permanent credential that works everywhere. The result can be returned to the caller while the secret remains outside the caller's general configuration.

This is useful for deployment hooks, notification delivery, provider calls, and service actions that should be revocable without rebuilding every workload.

AI-assisted work without secret context

An AI workflow can choose an allowed operation and ask an authorised machine to perform it. The model sees the task, the available operation, and the result. It does not need the provider token, database password, or signing key.

That is a narrower and more useful boundary than asking a model to edit a plaintext .env file and hoping the value does not appear in the next transcript.

Rotation and evidence

References can remain stable while the value behind them changes. A deployment does not need to be rewritten only because a provider credential rotated. The access decision and use can remain part of the event and audit trail.

Need the implementation details?

Read the Tasmanian Cloud documentation ↗