Identity for AI agents: the missing control
Published · Updated
8 min read
Agents act with real authority but often inherit a human's credentials. Here is what a first-class agent identity looks like.
Borrowed credentials do not scale
The fastest way to give an agent access is to hand it an API key or a session cookie. It works, and it is also indistinguishable from credential theft: the agent can do everything the human can, forever, with no separate record of who acted.
Once several agents share one human identity, logs stop answering the only question that matters after an incident: which actor did this?
What a real agent identity needs
An agent should have its own identifier, rooted in the human who delegated to it. Its authority should be scoped to specific actions on specific resources, time-boxed, and revocable in one click without disturbing the human's own access.
Delegation should only ever narrow. An agent cannot sub-delegate authority it was never granted, so a chain of agents can never end up with more power than the person at the top of it.
Rooted in a passkey
Authiun anchors every agent grant to a passkey held by a person. Revoking the grant, or the passkey, ends the agent's authority immediately across every connected service.
The result is an audit trail that reads as it should: this human authorised this agent to do this action, at this time, under this scope.