A secure vault your AI can use without ever exposing a secret

An agent that can act on your behalf needs credentials. An agent that holds credentials in its context is a credential leak waiting for the right prompt.

Heyno Passwords separates the two.

The problem with agents and secrets

Anything in a model’s context can, under adversarial pressure, come back out of it. Prompt injection through a document, a message, or a web page is a live attack class, and the assumption has to be that context is extractable.

The mitigation is not a better prompt. It is never putting the secret in the context.

Use without disclosure

Credentials live in an encrypted vault the model cannot read. When a task requires one, the model requests an action: sign in here, authorize this. A separate execution layer performs it with the secret. The model receives the result, never the value.

Scoped, expiring grants

Every use is a scoped, expiring grant tied to a specific task, not standing access. A grant that is not consumed expires on its own.

What is logged

Every access is recorded with which credential, which task requested it, and what it was used for. The log is readable by the workspace owner and is not editable by the assistant.

If something goes wrong

Rotation is a first-class operation. A credential can be rotated from inside the vault and every task depending on it picks up the new value without re-authorization.