Identity Management for Agentic AI [pdf] (2025)
openid.net
openid.net
https://auth0.com/docs/ai-agents-mcp/cross-app-access
https://datatracker.ietf.org/doc/draft-ietf-oauth-identity-a...
https://x401.proof.com/spec/latest/#abstract
in a nutshell:
* a website that wants to authorize who you are (say, to book a flight or sign a waiver for go kart rental)
* the endpoint returns 401 and defines in a header what info it needs about you (over 18? you're actually John Doe? etc.)
on the proof side specifically, we're putting IAL2 verification in front of this https://pages.nist.gov/800-63-3-Implementation-Resources/63A...
pretty cool stuff, its early days but its a strong way to ensure there's a human authorizing sensitive actions an agent is taking on your behalf
Authorizations are what are granted to an authenticated identity, typically with a specified scope and duration.
when I say `authorize who you are` I mean to say that you're saying both "hello I am in fact john doe" and "john doe the human is also saying this is ok to do".
I think this is interesting in the lens of Muse, GrokBot, Dots, OpenClaw, etc; if my agent wanted to rent a car on my behalf, it would forcibly have to get approval from me to do so
Authorization is the process of determining if someone can do something.
That you’re conflating multiple distinct concerns doesn’t build confidence.
I've only ever seen authorization header containing credentials (i.e. authentication, who you are) instead of authorization (what you can do).
Also everyone returns 401 when unauthorized (i.e. can't do a thing), instead of 403 (forbidden, i.e. can't do the thing). When 401 should probably be "unauthenticated" (we don't know who you are, so we can't authorize you).
Always messes with my head a bit.
- OSS (https://github.com/awebai/aweb/tree/main/awid)
- Trust rooted in the DNS.
- Multiple registries supported.
- did, verifiable stable identities.
- Certificate-based teams.
It is being a very interesting project so far. It's at the core of my https://aweb.ai which enables agents to communicate globally, and it makes it possible to build really simple agent-first tools where auth is just a matter of validating a certificate.
You could split use cases by scope. One is internal organization and IAM cases, which makes it easy for you to use a centralized identity provider (IdP) to issue some sort of token grant access to an app.
Use cases where you don't want to fix an IdP, like payments or verifying the someone's credentials before giving access on the open web, get more involved. You could have multiple valid IdPs, reputation / claims linked to IDs (since being on the IdP itself is no longer sufficient for an app to let the actor through), inheriting authority through multiple hops (giving your authorization to someone), enforcing narrowing permissions for those use cases, handling revocations. And then there's also the matter of being compatible with other emerging standards that are coming up (eg. A2P for agent payment authorization).
We settled on the core primitives of W3C DIDs for identity, Verifiable Credentials for delegations, and StatusList bitstrings for revocations as the "minimal set of ingredients" that address all the above. They are all open standards. DIDs are particularly cool because you can have them bound to passkeys, web domains through existing PKI, and even ledgers.
We did need to add a few specs of our own on top to address how it would work with use cases like MCP (for replacing OAuth flows, which have had messy implementations on coding agents since inception.) Our MCP example [0] also tries to make it incrementally adoptable and work alongside OAuth, so for example an operator could add it as a wrapper today and start collecting logs with identity claims for each tool call for auditability. It's a work in progress so hearing feedback on what seems confusing / hard to understand would be helpful.
I think this is the cursed part of the mission.
What are we actually arguing for here? Something like a limited liability corporation? "Agent-native identity" reads to me the same way that "unaccountable" does.
Creating a "MyAiRobot" account in GitHub and then rejecting ownership over that account is where this starts to turn into a problem. We need to make a human responsible for these things at all times. Any human will do, as long as they agree to the consequences of this ownership. The overall goal here is for the responsible party to be strongly incentivized to encourage good behavior down the org chart. In much the same way the compliance team at a bank tends to pull strings in such in a way that the operations staff won't piss off the FDIC.
We already have extensive means to pin AI activity to specific user principals. We do not need an entire revolution in authn protocols. There are more than enough existing tools to solve this problem.
I do enjoy asking people what they mean when they say they want to build an agent-native platform, because that work is rarely what they actually want to do.
Agents are not identities, they need to be granted permission to accomplish tasks. The identity is the person that grants the permission.
If no identity is required then this is simple machine-machine API keys.
The rapid rise of AI agents presents urgent challenges in authentication, authorization, and identity management. Current agent-centric protocols (like MCP) highlight the demand for clarified best practices in authentication and authorization. Looking ahead, ambitions for highly autonomous agents raise complex long-term questions regarding scalable access control, agent-centric identities, AI workload differentiation, and delegated authority. This whitepaper is for stakeholders at the intersection of AI agents and access management. It outlines the resources already available for securing today’s agents and presents a strategic agenda to address the foundational authentication, authorization, and identity problems pivotal for tomorrow’s widespread autonomous systems.