Speaker
Description
AI agents are usually authorized as if they were ordinary applications. This leaves an important gap. Existing systems can identify an application and grant access to an account, but they do not adequately describe what one agent instance may do for one task. They also provide limited ways to control how authority is passed to other agents.
This talk examines that missing authorization layer through GNAP, an IETF standard for delegated authorization. GNAP supports structured access requests, client instance keys, proof of possession, continuation and revocation. These mechanisms are already implemented in open source payment infrastructure through Rafiki and Open Payments. The talk considers how GNAP could support AI agents without replacing existing APIs or credentials. An agent receives limited authority to call a mediation service. That service retains the underlying credential and checks each requested operation before it reaches the downstream system. The talk separates what GNAP already provides from what still needs to be designed for agents. This includes task specific permissions, delegation between parent and child agents, and shared limits across multiple agents.
Summary
Most authorization systems grant access to users or applications, not to individual AI agents performing specific tasks. This talk examines whether GNAP can provide limited, revocable and auditable authority for agents. It also identifies what GNAP already provides and what must still be designed for safe delegation between agents.
Speaker Bio
Golang Engineer at Interledger Foundation
| Speaker LinkedIn / Socials | https://www.linkedin.com/in/daev-mithran-30b314158 |
|---|---|
| Difficulty | Intermediate |