JWT Profile for OAuth 2.0 Access Tokens
RFC 9068 · RFC9068 · JWT access token profile · at+jwt
By Boston Identity · Explained · 7 min read
The short answer
RFC 9068 is the IETF's standard layout for an OAuth access token written as a signed JSON Web Token. It fixes the token type, the claims every token carries, and the checks an API must run before trusting it.
- What it is
- A standard
- Defined by
- IETF, RFC 9068 (Standards Track)
- Who it's for
- Teams running OAuth servers and the APIs behind them
- In our briefings
- 320by month, last 12
What is JWT Profile for OAuth 2.0 Access Tokens?
OAuth 2.0, the framework for issuing access tokens, does not set any format for them.§1 Many vendors chose the JSON Web Token (JWT) format, a signed bundle of named fields called claims, so that an API could check a token itself without asking the server that issued it.§1 Their tokens did much the same job but used different claim names and syntax.§1
RFC 9068 offers one shared profile as an alternative to those private layouts.§1 It sets the mandatory and optional claims, says how a request's parameters shape the token, and tells the receiving API how to validate it.§1 The aim is that an authorization server from one vendor and a resource server (the API holding the data) from another can issue and accept the same tokens.Abstract It is a published IETF standard, issued in October 2021.S1
Not the same as the JWT bearer grant of RFC 7523, which uses a JWT to ask for an access token; RFC 9068 describes how to write the access token itself as a JWT.§1 It is also not an OpenID Connect ID token, the login token with a very similar layout, and the profile's explicit type exists so an API can tell the two apart.§5
How does JWT Profile for OAuth 2.0 Access Tokens work?
Three parties take part: the client (the app or agent that wants access), the authorization server that issues tokens, and the resource server, the API the token is for.
Figure steps
- The client asks for a token
- The authorization server picks the audience
- The authorization server signs the token
- The client passes the token on unread
- The resource server validates it
- The resource server decides
The authorization server writes the API's name into the token's audience and signs it, and the API checks type, issuer, audience, signature and expiry against the server's published keys before it decides.
Drawn from RFC 9068, sections 2 to 4
Swipe for the full figure →
The client asks for a token
The client uses any OAuth grant that ends in an access token.§3 Its request can name the target API in a resource parameter and list the permissions it wants in a scope parameter.§3 The profile gives the client no way to ask for particular claims; the server decides those from the client, scope and resource.§2.2.2
The authorization server picks the audience
The audience, held in the aud claim, names the API the token is for. If the request carried a resource parameter, aud should match it.§3 If not, the server must use a default resource and should infer it from the scope; when scopes point at different defaults, it should reject the request with invalid_scope.§3 It must not issue a token whose grant would be ambiguous.§3
The authorization server signs the token
The token must be signed, never with the "none" algorithm, and asymmetric keys are recommended.§2.1 Every conforming server must support the RS256 algorithm.§2.1 The header must declare the type application/at+jwt, which should be written as at+jwt.§2.1 Seven claims are required: iss (issuer), exp (expiry), aud, sub (subject), client_id, iat (issue time) and jti (a unique token id).§2.2 Every scope string in the token must mean something for the APIs named in aud.§2.2.3 Scope, roles, groups, entitlements and login details such as acr are optional or conditional.S1
The client passes the token on unread
The client presents the token to the resource server. It must not inspect the token's content, because the servers may change the format at any time, for example back to opaque tokens.§6 The profile does not tie the token to the Bearer scheme; other authentication schemes remain possible.§4
The resource server validates it
The resource server must reject any token whose typ is not at+jwt or application/at+jwt.§4 It decrypts the token if encryption was agreed.§4 It must then check that iss exactly matches the expected issuer, that aud names this resource server, and that the signature verifies with the authorization server's keys, rejecting "none".§4 The issuer should publish those keys and its issuer value in its OAuth metadata, through jwks_uri and issuer.§4 Last, the current time must be before exp, with a few minutes' leeway allowed for clock skew.§4
Why does JWT Profile for OAuth 2.0 Access Tokens matter now?
One IETF draft shows a newer token design built on top of the profile. That draft defines selective-disclosure access tokens, where the client reveals some claims to an API and hides others. It keeps the token an RFC 9068 JWT with type at+jwt.1 A resource server that does not implement the draft ignores the new fields and sees only the cleartext claims.1 Its authors kept at+jwt on purpose, against RFC 9901's advice of a distinct type, because RFC 9068 tells servers to reject any other.1
That draft's motivating case is an agent calling several APIs on a user's behalf.2 What it offers the agent is privacy per request, not delegation: the act claim stays visible, and the client still receives every value.2 It also requires its tokens to be bound to a key with DPoP.2 We read this as a sign that, in this draft at least, an agent's token is an RFC 9068 token with more attached, so the base profile's checks still decide what gets through.
RFC 9068's own warning about sub now also appears in the security guidance it pointed to. RFC 9700, the OAuth security best practice, repeats that sub should name the user when one is involved and the client when none is.3 A malicious client that can pick its own client_id may pick a user's identifier and, if the API cannot tell the two kinds of token apart, reach that user's resources.3 Where client and user identifiers share a namespace, RFC 9700 says authorization servers should not let clients influence them, and must otherwise give APIs another way to tell the two apart.3
Where does it fit?
| Compared with | How JWT Profile for OAuth 2.0 Access Tokens differs |
|---|---|
| Resource Indicators for OAuth 2.0 | Resource Indicators (RFC 8707) defines the resource parameter a client uses to name its target API. RFC 9068 consumes it: the token's aud claim should carry the same value.§3 |
| OAuth 2.0 Authorization Server Metadata | OAuth Authorization Server Metadata (RFC 8414) is the published document listing a server's endpoints. RFC 9068 says issuers should use it to advertise their signing keys and issuer value, so APIs know what to check against.§4 |
| OAuth Token Exchange | Token Exchange (RFC 8693) trades one token for another. RFC 9068 takes its client_id and scope claims from it,S1 and says the login details in a token stay fixed across such exchanges.§2.2.1 |
| DPoP | DPoP is a separate mechanism for binding a token to a key the client holds. RFC 9068 does not require it but leaves room for schemes other than Bearer,§4 and the selective-disclosure draft sends its tokens under DPoP.1 |
What goes wrong?
- Token swaps between login and access. An OpenID Connect ID token looks almost the same, so an API that skips the type check can accept one as an access token.§5 The required at+jwt type is the defence.§4
- A client posing as a user. If a server uses client_id as sub for machine grants and lets clients pick their own client_id, a malicious client can choose a high-privilege user's sub.§5 RFC 9700 carries the same warning forward.3
- Tokens replayed across APIs. Servers must give each distinct resource its own aud value to prevent cross-JWT confusion, where a token meant for one API is accepted by another.§5
- Ambiguous scopes. When a request names several resources, each scope in the token should map to exactly one of them, and the profile leaves the method to each deployment.§5
- Any key can sign. An API cannot tell which published key is meant for access tokens, so an attacker who steals any one of them can mint tokens it will accept.§5
- Data on show. Unencrypted tokens can be read by the client and sometimes the end user, and a stable sub lets APIs correlate a person's activity.§6
What should a security leader ask?
- 01
Which of our APIs reject a token whose typ header is not at+jwt, and would any accept an OpenID Connect ID token instead?
- 02
Can any client, including an AI agent we register, choose its own client_id, and could that value collide with a user's sub?
- 03
Does every API we run have its own audience value, and do we test that a token for one is refused by another?
- 04
How many signing keys does our authorization server publish, and who can reach the private half of each?
- 05
Which personal attributes do we put in unencrypted access tokens, and which clients can read them?
Sources
- S1
JSON Web Token (JWT) Profile for OAuth 2.0 Access Tokens (RFC 9068)IETF · Standards Track RFC, October 2021
›Evidence: 3 claims from 2 briefings
- 1BI-FIND-2026-0458 · A draft token hides some claims, works with old servers unchanged, and leaves agent delegation claims visibleThe token stays an RFC 9068 JWT, typ at+jwt, sent as today under the DPoP scheme; what changes sits beside it. The authorization server returns a disclosures array next to access_token, and the client sends any subset in SD-JWT-Disclosures, a Structured Fields list in which order matters, with an optional key-binding JWT in SD-JWT-Key-Binding. Only top-level claims may be hidden, and iss, exp, aud, sub, client_id, iat, jti and cnf, plus scope and nbf when present, stay in the clear. A resource server that implements the profile validates the token as before and then reassembles and verifies the SD-JWT; one that does not 'ignores both fields' and 'sees only the cleartext claims', and 'a resource server MUST NOT read an undisclosed claim as permission'. The draft keeps at+jwt on purpose, against RFC 9901's advice of a distinct type, because RFC 9068 tells servers to reject any other.
- 2BI-FIND-2026-0459 · A draft token hides some claims, works with old servers unchanged, and leaves agent delegation claims visibleThe motivating case is an agent calling several resource servers on a user's behalf, and the example client is 'agent-7f3a', but what the profile offers an agent is attribute privacy per request, not delegation. The act claim and everything nested in it must remain in the clear, there is no rule for forwarding disclosures to a next hop, and the client 'does receive every value'; where it must not, 'encryption remains the only answer'. Unlinkability is not claimed, since jti, sub and the digests repeat on every call. Tokens 'MUST be DPoP-bound', the key-binding JWT exists to detect an intermediary stripping disclosures, and both fields are to be kept out of logs and caches like the Authorization header. The appendix's own open questions are whether that key-binding JWT is redundant beside DPoP, whether keeping at+jwt is acceptable, and how a server asks for a disclosure it needs: 'no existing error code fits valid Token, required claim not disclosed'.
- 3BI-FIND-2025-0354 · RFC 9700 bans OAuth's password grant and loose redirect matching; binding access tokens stays a SHOULDSection 4.15 warns that a resource server can mistake a client for a user. Under the JWT access token profile (RFC 9068), sub should name the resource owner when a user is involved, and the client application under grants with no user, such as client credentials. So if a client can choose its own client_id at registration, a malicious client may pick a user's identifier and, if the resource server cannot tell the two kinds of token apart, reach that user's resources. Where client IDs and user identifiers share a namespace, authorization servers should not let clients influence their client_id or any claim that could be confused with a real user's. Where that cannot be avoided, they must give resource servers some other way to distinguish the two.
›In our briefings
What our briefings say about JWT Profile for OAuth 2.0 Access Tokens itself, with the most direct claims first.
The token stays an RFC 9068 JWT, typ at+jwt, sent as today under the DPoP scheme; what changes sits beside it. The authorization server returns a disclosures array next to access_token, and the client sends any subset in SD-JWT-Disclosures, a Structured Fields list in which order matters, with an optional key-binding JWT in SD-JWT-Key-Binding. Only top-level claims may be hidden, and iss, exp, aud, sub, client_id, iat, jti and cnf, plus scope and nbf when present, stay in the clear. A resource server that implements the profile validates the token as before and then reassembles and verifies the SD-JWT; one that does not 'ignores both fields' and 'sees only the cleartext claims', and 'a resource server MUST NOT read an undisclosed claim as permission'. The draft keeps at+jwt on purpose, against RFC 9901's advice of a distinct type, because RFC 9068 tells servers to reject any other.
A draft token hides some claims, works with old servers unchanged, and leaves agent delegation claims visibleBI-FIND-2026-0458
When no user is involved, as in the client credentials grant, RFC 9068 says sub should identify the client application. Its security section warns where that can lead. If an authorization server uses the client_id as sub, it should stop clients from registering an arbitrary client_id, because a malicious client could pick the sub of a high-privilege resource owner and confuse any authorization logic on the resource server that relies on sub. The warning is written with a lowercase "should".
RFC 9068 makes resource servers check a JWT access token's type and audience, and warns a client could pose as a userBI-FIND-2021-0004
Registrars issue OAuth JWT access tokens (RFC 9068) whose subject is the agent's identifier, and each must be bound to a key the agent holds: "Bearer-only tokens without cnf MUST NOT be issued." RFC 9421 HTTP Message Signatures are the proof that every issuer and relying party must support, and mutual TLS is optional. The author's own services already work this way: the 1id.com and MailPal.com APIs accept a token only with a message signature by the enrolled key, and mutual TLS is not offered.
Agent registry draft lets another identity block an agent's key changes, not switch it offBI-FIND-2026-1725
›Cited in 32 briefings
Every briefing that discusses this term, newest first.
- An agent audit draft can now name the token issuer, subject and client behind a decision
Draft · · IETF Datatracker · Sankalp Gilda · Oct 16, 2026
- Agent registry draft lets another identity block an agent's key changes, not switch it off
Draft · · IETF Datatracker · Christopher Drake, 1id.com · Oct 27, 2026
- A disabled agent's old tokens still pass every check; a DNS draft now asks receivers to time how long
Draft · · IETF Datatracker · Sriram Shankarlingam, HexaEight · Oct 20, 2026
- Asor's agent-chain draft now says a chain keeps verifying after its principal leaves, until expiry or revocation
Draft · · IETF Datatracker · Rafael Asor · Oct 28, 2026
- An OAuth draft lets the API, not the agent, describe the approval, and now name the key the token is bound to
Draft · · IETF Datatracker · Yaroslav Rosomakho et al. · Jun 25, 2026
- Raut's revised agent draft keeps token exchange out of the model's reach, and concedes the token carries no evidence of it
Draft · · IETF Datatracker · Ashay Raut · May 22, 2026
- Sweeney's delegation draft concedes its server is trusted with users' credentials, and drops an interop claim
Draft · · IETF Datatracker · Kieran Sweeney · Oct 13, 2026
- Cross-App Access draft warns that one stolen sign-in token could reach every app the identity provider permits
Draft · · IETF Datatracker · Aaron Parecki et al. · May 21, 2026
- Global Token Revocation must end a user's refresh tokens; their access tokens can outlive the call
Draft · · IETF Datatracker · Aaron Parecki, Okta · Feb 25, 2026
- MCP's November 2025 authorization spec puts metadata documents ahead of dynamic client registration
Article · · Model Context Protocol · Nov 25, 2025
- Drake's agent registry names who may issue an agent's token, and its answer has no signature of its own
Draft · · IETF Datatracker · Christopher Drake, 1id.com · Sep 25, 2026
- An individual IETF draft takes a workload's role and group out of its name and puts them in separate metadata
Draft · · IETF Datatracker · Shlomo Zalman Heigh et al. · Oct 16, 2026
- An OAuth agent profile has the token server block any widening of scope, and leaves revoking the chain optional
Draft · · IETF Datatracker · Sanjeev Kumar, Orchestrum Technologies LLP · Aug 30, 2026
- In Huawei's reworked OAuth draft, evidence signed by the token's subject can narrow access but never widen it
Draft · · IETF Datatracker · Cheng-Kang Chu et al. · Sep 1, 2026
- TPX replaces an app's embedded AI provider key with a capped, metered grant the user approves at their provider
Article · · tokenpony.dev · Alex Polvi · Jul 20, 2026
- A draft token hides some claims, works with old servers unchanged, and leaves agent delegation claims visible
Draft · · IETF Datatracker · Su-hyeon Forten Lee · Oct 3, 2026
- Batch delegation assigns each permission to an agent, then filters at token exchange
Draft · · IETF Datatracker · Yuan Ni et al. · Jul 3, 2026
- Atlassian's first draft for offline delegation banned bearer handles; revision 05 just recommends sender binding
Draft · · IETF Datatracker · Larry Zhu et al. · May 22, 2026
- Three vendors put MCP controls in three different places in May, and only Auth0's was GA
Release tracker · · 10 vendors' release notes · May 2026
- A 'decentralized' agent identity draft blocks risky calls when its Registry is down, and did:aip is taken
Draft · · IETF Datatracker · Paras Singla · Apr 17, 2026
- OAuth 2.1's October draft bans plain PKCE and adds browser-app rules; 2.1-only clients can fail on older servers
Draft · · IETF Datatracker · Dick Hardt et al. · Mar 2, 2026
- Raut's January draft put AI agents in two new transaction-token claims; by April he used OAuth's act and sub
Draft · · IETF Datatracker · Ashay Raut · Jan 20, 2026
- The draft replacing RFC 8725 requires rejecting malformed JWTs but only recommends denial-of-service limits
Draft · · IETF datatracker · Yaron Sheffer et al. · Jan 12, 2026
- A WSO2 draft puts the AI agent in OAuth consent and the access token, but not how the agent gets its credential
Draft · · IETF Datatracker · Thilina Shashimal Senarath et al. · Aug 26, 2025
- The July 2025 Cross App Access draft lets the IdP decide app-to-app access and treats agents as plain clients
Draft · · IETF Datatracker · Aaron Parecki et al. · Jul 2, 2025
- MCP's June 2025 authorization spec: a server accepts only tokens issued for it and passes none upstream
Article · · Model Context Protocol project · Jun 18, 2025
- RFC 9700 bans OAuth's password grant and loose redirect matching; binding access tokens stays a SHOULD
Article · · IETF (RFC Editor) · Torsten Lodderstedt et al. · Jan 2025
- In May 2023, Parecki described OAuth 2.1 as consolidation and DPoP as token protection
Video · · Identiverse - A CRA Resource · Jun 17, 2024
- RFC 9470 lets an API demand a fresher or stronger sign-in, with no shared scale for 'stronger'
Article · · RFC Editor · Vittorio Bertocci et al. · Sep 8, 2023
- RFC 9068 makes resource servers check a JWT access token's type and audience, and warns a client could pose as a user
Article · · RFC Editor · Vittorio Bertocci, Okta · Oct 1, 2021
- RFC 7519 makes "none" mandatory to implement and leaves refusing an unwanted algorithm as a SHOULD
Article · · RFC Editor · Michael B. Jones et al. · May 1, 2015
- RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023
Article · · RFC Editor · D. Hardt et al. · Oct 2012
Related terms
What appears alongside JWT Profile for OAuth 2.0 Access Tokens in the same briefings.
Resource Indicators for OAuth 2.0OAuth 2.0 Authorization Server MetadataDPoPPKCEOAuth Identity and Authorization Chaining Across DomainsRich Authorization RequestsOAuth Token ExchangeOAuth 2.0 Security Best Current Practice
Was this explainer useful?
Share this explainer
Every finding, forecast and open question on this site has a permanent ID. If we reword one, its ID stays the same, so a citation keeps pointing to the same claim.
Working on an identity problem that involves JWT Profile for OAuth 2.0 Access Tokens? Talk to Boston Identity