IETF email OAuth draft drops refresh keys and details replay risk
News
Published 20 Nov 2026
Updated 20 Nov 2026
7 min read
Summarize with

The IETF MAILMAINT working group's revision 08 of OAuth Profile for Native Clients of Open Protocols removes the proposed refresh_key mechanism introduced in revision 07. It changes how refresh-token retries would work and explicitly analyzes a residual risk: a stolen refresh token can retain access across rotations without triggering superseded-token detection.
This is an active working-group Internet-Draft, not a final standard or production guidance. It is not a current Gmail, Yahoo, Microsoft, or other provider requirement. Datatracker lists WG Document and I-D Exists, with Informational as the intended final status. It has not been approved as an RFC. There is no rollout, enforcement date, or deployment deadline.
The revision was published and announced on November 19, 2026 UTC. The official announcement is timestamped 02:28 UTC, with receipt at 02:28:38 UTC; the document cover is dated November 18.
Revision 08 removes the auxiliary refresh key
In revision 07, the proposed refresh_key parameter was an extra random value associated with a refresh request. A prototype following that revision would persist it before sending, reuse it on retries, and let the server compare it with the recorded value to distinguish an eligible retry.
Revision 08 removes that request parameter, the client persistence and reuse rules, the server key-matching mechanism, and the requested IANA registration. Our earlier refresh-key report covered revision 07's addition. Revision 08 has now superseded that proposed mechanism; neither revision should be read as a universally deployed requirement.
|
|
|
|---|---|---|
Auxiliary key | Proposed | Removed |
Retry check | Matching key | Old token valid |
Replacement | Same refresh token | Same refresh token |
Replay discussion | Key-based detection | Residual risk explicit |
Proposed behavior in revisions 07 and 08, not deployment status.
The removal changes both sides of a prototype: the client would stop relying on auxiliary key state, while the server would need a different rule for deciding when the previous refresh token becomes invalid. The remaining refresh token is still a credential, not a disposable request identifier.
The proposed overlap depends on evidence of receipt
Under revision 08, a server SHOULD NOT invalidate the previous refresh token immediately after issuing its replacement. Instead, it SHOULD keep the old token valid until it has evidence that the replacement response reached the client. This is a recommendation, not a universal MUST: immediate invalidation remains possible.
Observable use of the replacement access token would provide that evidence. So would presentation of the replacement refresh token at the token endpoint. The distinction matters because an authorization server cannot always observe access-token use, particularly with self-contained tokens validated locally or separate IMAP and SMTP servers.
Access-token evidence
Where use is observable, a server following the recommendation would end the overlap when the replacement access token is first used. Separate resource servers complicate that visibility.
Refresh-token evidence
Presentation of the replacement refresh token is visible at the token endpoint. It is the receipt evidence available to every authorization server in the proposed model.
The operational problem is a lost response. If a network interruption or operating-system suspension prevents the client from saving its replacement, immediate invalidation can leave it holding only an unusable token. Revision 08 proposes recovering through overlap. Revision 06 already had a more general grace-period concept; overlap itself is not new to this work.
For client retries, revision 08 proposes SHOULD retry with the same refresh token after a lost-response failure, and MUST NOT discard it except after invalid_grant or invalid_client. Those are draft implementation rules, not instructions to alter an existing provider integration.
The same-replacement safeguard is retained
While the previous refresh token remains valid, revision 08 would require a retry to return the same replacement refresh token. It would prohibit issuing a separate replacement that starts an independent token chain. This safeguard is retained from revision 07, although the condition for accepting the retry changes.
Once the old token is superseded, the proposed response to presenting it is invalid_grant and revocation of the entire authorization, including the current refresh token and all associated access tokens. Revision 08 also retains replay history, checks for superseded tokens before validating other request parameters, and atomic processing of requests for the same authorization.
These retained protections matter because concurrent requests must not create independent successors or receive a successful retry response after supersession. A prototype review should distinguish a retry accepted during overlap from use after confirmed supersession. The new issue is how much confidence a server can draw from token rotation while the old token remains acceptable.
The draft explicitly describes residual replay risk
The expanded revision 08 analysis explains the limitation for refresh tokens without sender constraints. During overlap, the server cannot distinguish a legitimate lost-response retry from a request carrying a stolen copy of the same token. Both can receive the same replacement refresh token.
A correctly timed attacker can keep obtaining those replacements across successive rotations without presenting a superseded token. The analysis says access can continue indefinitely, meaning for as long as the authorization remains valid. It does not require repeated theft of the client's stored credentials.
Rotation has a detection boundary
This is a risk analyzed in a proposed design. It is not evidence of active exploitation, a provider breach, or a defect in every deployed OAuth implementation. Rotation does not detect every stolen-token use.
The discussion also explains why the risk is broader than winning a narrow race after the legitimate client refreshes. An attacker refreshing first and delaying use can avoid that narrow timing dependency. Closing overlap at first observable access-token use reduces exposure, but does not eliminate the described strategy.
My concern with describing rotation as automatic theft detection is this distinction: presenting an already superseded token is detectable, but keeping pace inside the proposed overlap does not necessarily produce that event.
Proposed mitigations require server-side judgment
Revision 08 proposes additional signals rather than claiming the overlap is risk-free. Under its guidance, servers SHOULD treat repeated presentations of previous tokens across successive rotations as a compromise signal and revoke the authorization. They MAY also consider differences between requests, including source network addresses.
The draft further proposes that servers SHOULD end overlap at the first observable use of the replacement access token. That advice depends on visibility: the token endpoint cannot infer every use at an independent resource server. These are proposed mitigations for implementers to evaluate, not immediate production configuration instructions.
DPoP does not become a universal fix through this revision. Its proofs operate over HTTP; DPoP-bound access tokens do not suddenly work in ordinary OAUTHBEARER IMAP or SMTP exchanges. General DPoP support was already discussed in earlier revisions, and revision 08 does not replace RFC 9700 as current production guidance.
Who should review the revision now
The directly affected audience is developers of native installed email clients and authorization servers following or prototyping this draft. Its scope includes JMAP, IMAP, SMTP, and POP, plus CalDAV and CardDAV. Web clients are explicitly excluded.
For a revision 07 prototype, I would focus the next review on the changed token request and lifecycle assumptions, then continue following MAILMAINT discussions before planning production changes. Removing an auxiliary parameter is straightforward; determining when an old token becomes superseded needs review across the authorization server and resource servers. Useful review tasks include:
- Request state. Compare removal of refresh_key with the prototype's stored retry state.
- Receipt signals. Establish which replacement-token uses the authorization server can observe.
- Replay testing. Review lost responses, concurrent refreshes, and repeated old-token presentations.
- Revocation scope. Check that superseded-token handling reaches the entire authorization.
Ordinary domain administrators need no SPF, DKIM, DMARC, or BIMI DNS change because of this draft. Existing senders should continue meeting their current authentication and provider requirements. OAuth account access and domain-level message authentication need separate investigation.
Suped is our DMARC reporting and email authentication platform. For ongoing sender checks, Suped's DMARC monitoring helps investigate sending sources and authentication results. Our email tester analyzes a real message. Neither checks an authorization server's refresh-token replay logic, and neither is a prerequisite for following this draft.
What to watch next
The next useful signal is how MAILMAINT develops the balance between recoverable retries and compromise detection, particularly where access-token use is invisible to the authorization server. Teams tracking the work should review subsequent revisions and working-group discussion before committing to a production design. There is no draft-driven deployment deadline for domain owners.
Frequency asked questions
No. Revision 08 removes refresh_key and its associated client, server, and IANA provisions. Our earlier report describes revision 07, whose proposed mechanism has been superseded.
No. During the proposed overlap, a valid old token would return the same replacement refresh token. Presenting it after supersession would require invalid_grant and authorization-wide revocation under the draft.
No. The draft analyzes residual replay risk in its proposed design. It does not announce a provider incident, implementation uptake, or evidence of active exploitation.
No. This unfinished OAuth profile introduces no SPF, DKIM, DMARC, or BIMI DNS action for domain owners. Current provider requirements continue independently.

