Jia, Meiling,

On #6 — on-device agents in the cloud OAuth ecosystem. It reads as a 
communication question, but I think it sits inside a security one, so let me 
take it a little wider than communication alone.

Start with what the endpoint itself can see — presence, posture, local 
standing. This is where Yossif's #11 comes in: a condition-bound key that 
disappears when the condition is no longer met, so it signs only while those 
conditions hold. For everything in that set, the endpoint needs no 
communication at all. The key stops signing in the TLS key agreement and access 
stops. Nothing to send.

For what the endpoint cannot see — the user withdrawing consent, the agent 
being distrusted — some external entity has to raise the signal. Usually that 
signal goes to the RP, not the endpoint, and CAEP is the standard way to carry 
it.

There can be cases where the signal has to reach the endpoint instead, and that 
is what #6 seems to be about. I'm not sure there's an established way. CAEP or 
a lighter version of it can do it. Or that’s what you are trying to cover?

But that path — a signal sent to the endpoint — is rarely needed. The endpoint 
terminates its mTLS on the conditions it already knows, and the external 
signals go to the RP. So there is rarely a need for the endpoint to talk back 
to the OAuth authority at all. I might be wrong though 😊.

draft-winmagic-wimse-condition-bounded-credentials carries the detail. Happy to 
go into it if useful.

Thi

From: 陈佳 <[email protected]>
Sent: Wednesday, July 22, 2026 4:37 AM
To: oauth <[email protected]>
Subject: [OAUTH-WG] Re: New Version Notification 
fordraft-chen-oauth-agent-authz-use-cases-01.txt

You don't often get email from 
[email protected]<mailto:[email protected]>. Learn why this is 
important<https://aka.ms/LearnAboutSenderIdentification>
CAUTION:This email originated from outside of the organization. Do not click 
links, open attachments or respond unless you recognize the sender and know 
that the content is safe.


> Jeff Lombardo wrote on Mon, 13 July 2026 10:50 UTC:

> "Here are some comments on the elements listed as Gaps..."

> "you identify that OAuth is not capable of understand an intent.

> True but is it the purpose of OAuth?"

> "Impractical Revocation... write operations could be perform

> through the usage of Transaction Tokens"

> "If tokens are cryptographically bound to a client key, they

> cannot be shared... See DPoP (RFC9449)"

> "This is a big misconception. OAuth is perfectly capable of

> handling claims... If you are looking for standards, you should

> look at OpenID Foundation AuthZEN"

> "there is progress in the right direction with ID JAG and

> Transaction Token Authorization Grant Profile for OAuth Identity

> and Authorization Chaining"



Thanks, Blake, Yossi, Jeff, and Thi, for the detailed discussion.

Your feedback is very helpful as we refine the draft.



I agree with the specific technical points raised — especially the

pointers to TXN tokens, DPoP, CIMD, ID-JAG, and AuthZEN. Each of

these is moving in the right direction.



However, I want to zoom out and look at the overall picture.



Even though multiple OAuth-related projects exist, they currently

lack a systematic connection. It is hard to see them forming a

complete authorization solution for agent security.



Let me illustrate with one concrete example:



- Transaction Tokens can trace the full operation chain — good.

- But what is their relationship with agent identity?

- What about the agent's dynamic, per-operation permissions?

- Can the existing building blocks answer this core question: an

  agent inherits permissions from the user, yet those permissions

  may change dynamically with each operation?



I believe we need to answer the following questions systematically,

often requiring cross-project collaboration:



1. Binding between agent identity and user identity — whom does the

   agent represent, and how is this binding expressed and verified

   during authorization?

2. Binding between agent identity and operation — which agent

   initiated each operation? How to prevent identity spoofing?

3. Traceability of agent actions — TXN tokens should serve as a

   post-hoc audit trail forming a complete "agent -> operation"

   trace tree. That tree does not exist today.

4. Non-repudiation of agent actions — once an operation is

   completed, can the agent or its principal deny it? We need

   cryptographic guarantees.

5. Permission inheritance — how does an agent inherit permissions

   from the user? What are the boundaries? Can inheritance be partial?

6. Local agent integration — how do on-device agents interact with

   the cloud-based OAuth ecosystem? This remains underexplored.

7. Intent propagation — the user's original intent should travel

   with each operation, so the executing party can verify

   consistency.



Jeff's references are valuable, but they feel like scattered puzzle

pieces. The challenge is not whether each piece exists, but whether

they can be assembled into a coherent picture. We need a systematic

solution, not isolated fixes for each sub-problem.



The goal of this draft should be to identify these systemic gaps

and provide a common reference framework for future cross-project

collaboration.



Looking forward to further discussion.



Best regards,

Jia Chen



E-mail:  [email protected]<mailto:[email protected]>




_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to