# What is SD-JWT?

> SD-JWT is a signed JSON Web Token in which the issuer replaces chosen claims with salted hashes, so the holder can show each verifier only the claims it picks and, optionally, prove it holds a bound key.

Explainer by Boston Identity, 2026-10-29. 7 min read. https://insights.bostonidentity.com/glossary/sd-jwt

- What it is: standard
- Defined by: IETF, RFC 9901 (Standards Track)
- Who it's for: Teams accepting credentials from wallets or agents

## What is SD-JWT?

A JSON Web Token (JWT) is a signed bundle of JSON (JavaScript Object Notation) claims, often used to describe a user [S1 §1]. When a server sends one straight to the site that uses it, the server can simply leave out claims at issuance [S1 §1]. SD-JWT, short for Selectively Disclosable JWT, serves a different model. An issuer gives a token with many claims to a holder, and the holder later shows it to several verifiers, each of which may need only some of the claims [S1 §1]. The RFC's own example is a token with an address and a birthdate: the holder shows the address to one verifier and the birthdate to another [S1 §1].

The method is salted hashes. For each claim that should be hideable, the issuer signs a digest in place of the value, and the holder sends the value with its random salt only to a verifier that should see it [S1 §1]. That verifier hashes what it received and checks that the digest is in the signed payload. The salt stops a verifier from guessing hidden values [S1 §1].

The Internet Engineering Task Force (IETF) published the specification as RFC 9901 on the Standards Track in November 2025 [S1]. It also defines optional key binding, in which the holder proves possession of a private key whose public key sits in the token, to stop the token being presented without the holder's consent [S1 §1].

Think of it like this: Think of a signed cover sheet listing a fingerprint for each of your index cards: you hand a clerk the sheet plus only the cards they need, and the clerk checks each card against its fingerprint. The comparison breaks down on tracking, because a clerk who keeps the signed sheet can show it to whoever issued it.

Not the same as: Not an anonymous credential: RFC 9901 says it does not meet those security properties, and colluding verifiers and issuers can recognise the same token whatever was disclosed [S1 §10.1].

## How does SD-JWT work?

Three parties take part: the issuer, which creates the SD-JWT; the holder, which receives and controls it and may be the user, their device and software, or both; and the verifier, which checks what it is shown [S1 §1.2].

[Figure: The issuer signs digests in place of claims, the holder sends only the disclosures it chooses, and the verifier hashes each one to find it in the signed token.]

1. **The issuer swaps claims for salted digests.** For each claim it chooses to make hideable, the issuer builds a small array of a random salt, the claim name and the value, and encodes it as text; that string is a disclosure [S1 §4.2.1]. It hashes each disclosure, with SHA-256 by default, and puts the digests in the payload in place of the claims [S1 §4.2.3]. It may add decoy digests to hide how many claims there are [S1 §4.2.5]. It then signs the JWT with its private key [S1 §4.1].
2. **The issuer hands everything to the holder.** The issuer sends the signed JWT followed by all the disclosures, each separated by a tilde (~) character [S1 §4]. It issues a plain SD-JWT, and a holder must reject one that arrives with key binding attached [S1 §7.2]. On receipt the holder must validate the token and check that its claims are acceptable [S1 §7.2].
3. **The holder chooses what to reveal.** For a presentation, the holder decides which disclosures to release, getting consent if necessary; how consent is obtained is outside the RFC [S1 §7.2]. It may send none, some or all of them, and must not reveal the salts of claims it holds back [S1 §4]. A nested disclosure is useless without the disclosure that contains it, so the holder must send both [S1 §4.2.6].
4. **The holder signs a Key Binding JWT, if required.** When key binding is required, the holder creates a Key Binding JWT (KB-JWT) tied to the token and appends it after the last tilde [S1 §4, §7.2]. The KB-JWT must carry its issue time (iat), the intended verifier (aud), a nonce for freshness, and sd_hash, a hash over the signed JWT and the chosen disclosures [S1 §4.3]. The holder signs it with the private key matching the public key in the token [S1 §4.3.2].
5. **The verifier checks the signature and every disclosure.** The verifier first decides from its own policy whether key binding is required, and rejects a presentation without one when it is [S1 §7.3]. It checks the issuer's signature and key and refuses the none algorithm [S1 §7.1]. It hashes each disclosure, finds the matching digest, and puts that claim back in place; digests with no disclosure are ignored [S1 §7.1]. A disclosure that matches no digest, or a digest that appears twice, gets the token rejected, as does a missing validity claim it requires [S1 §7.1].
6. **The verifier checks the key binding.** If key binding is required, the verifier takes the holder's public key from the token and validates the KB-JWT signature with it [S1 §7.3]. It checks that typ is kb+jwt, that iat falls in an acceptable window, and that nonce and aud match this transaction and this verifier [S1 §7.3]. It recomputes sd_hash over what it received, so a changed set of disclosures fails [S1 §7.3].

## Why does SD-JWT matter now?

SD-JWT has become the token under several agent payment designs. Google's AP2 payment protocol authorizes an agent's purchase with a checkout mandate and a payment mandate, and both travel as SD-JWTs [BI-FIND-2026-0414](https://insights.bostonidentity.com/ref/BI-FIND-2026-0414). Verifiable Intent chains three signed SD-JWT layers, from card issuer to user to agent [BI-FIND-2026-0477](https://insights.bostonidentity.com/ref/BI-FIND-2026-0477). EMVCo's agent-payment framework presents Verifiable Intent as one realisation of its three layers and specifies SD-JWT VC, the IETF's verifiable-credential format built on SD-JWT, in section 6.7 [BI-FIND-2026-0440](https://insights.bostonidentity.com/ref/BI-FIND-2026-0440).

The standards on top are still moving. SD-JWT VC is in evaluation by the Internet Engineering Steering Group (IESG), and we forecast it is published as an RFC by 31 October 2027 [BI-CALL-2026-0077](https://insights.bostonidentity.com/ref/BI-CALL-2026-0077). The chairs of the IETF's OAuth working group, which maintains the OAuth authorization standards, opened a call for adoption of the Delegate SD-JWT draft on 28 September 2026; all fifteen replies we counted on 6 October supported it [BI-FIND-2026-0447](https://insights.bostonidentity.com/ref/BI-FIND-2026-0447).

RFC 9901 itself never mentions agents or delegation [BI-FIND-2025-0143](https://insights.bostonidentity.com/ref/BI-FIND-2025-0143). Key binding is optional and each verifier decides whether to require it, which we read as making replay protection a setting each verifier has to get right [BI-FIND-2025-0144](https://insights.bostonidentity.com/ref/BI-FIND-2025-0144).

## Where does it fit?

- Compared with Key Binding JWT: The Key Binding JWT is one optional part of an SD-JWT presentation: the holder's signed proof of key possession, covering a nonce, an audience and a hash of the SD-JWT as sent [S1 §3.3]. SD-JWT is the whole format, which also works without it.
- Compared with Delegate SD-JWT: Delegate SD-JWT lets the key-binding slot hold a selective-disclosure token of its own that can name the next holder's key, so a verifier can trace hand-offs back to the issuer [BI-FIND-2026-0444](https://insights.bostonidentity.com/ref/BI-FIND-2026-0444). RFC 9901 has no such chain [BI-FIND-2025-0143](https://insights.bostonidentity.com/ref/BI-FIND-2025-0143).
- Compared with Token Status List: Token Status List handles revocation: the issuer publishes one signed, compressed bit array, and a verifier reads the bits for the token in front of it [BI-FIND-2026-0632](https://insights.bostonidentity.com/ref/BI-FIND-2026-0632). RFC 9901 defines no status or revocation check of its own [BI-FIND-2025-0145](https://insights.bostonidentity.com/ref/BI-FIND-2025-0145).
- Compared with OpenID for Verifiable Presentations: OpenID for Verifiable Presentations (OpenID4VP) is a protocol for presenting credentials, not a token format; the OpenID Foundation's presentation tests cover both SD-JWT and mdoc credentials [BI-FIND-2025-0230](https://insights.bostonidentity.com/ref/BI-FIND-2025-0230).

## What goes wrong?

- Without key binding, anyone who gets the token can replay it, and a malicious verifier can present it to another verifier that does not require binding [S1 §9.5]. Verifiers must decide whether binding is required before verifying, and must not base that on whether a KB-JWT arrived, because an attacker could strip it off [S1 §9.5].
- In a bare SD-JWT the set of disclosures is not integrity protected, so the token can be changed by adding or removing disclosures and still be valid [S1 §9.10]. Any party holding it can forward it, with fewer disclosures, to a receiver that does not enforce key binding [S1 §9.9]. As we read the format, nothing in it names who forwarded it [BI-FIND-2025-0143](https://insights.bostonidentity.com/ref/BI-FIND-2025-0143).
- Unlinkability between issuer and verifier cannot be achieved when the verifier is careless, colluding, compromised or coerced, because it can forward the issuer-signed token to the issuer [S1 §10.1]. Batch issuance, with new key-binding keys and salts for each copy, can stop verifiers linking presentations [S1 §10.1].
- An issuer must not make content that is critical to the token's validity selectively disclosable, and verifiers must make sure every validity claim they need is present [S1 §9.7].
- Hidden values are only as safe as their salts, which must be random and infeasible to guess; the recommended minimum is 128 bits [S1 §9.3].
- RFC 9901 defines no status or revocation check for an issued token [BI-FIND-2025-0145](https://insights.bostonidentity.com/ref/BI-FIND-2025-0145).

## What should a security leader ask?

1. Which of our verifiers require key binding for SD-JWTs, and is that decision fixed in policy before a token is read?
2. When an agent presents an SD-JWT for a user, how does our verifier tell the agent from the user, given that the format names no forwarder?
3. How do we learn that a credential we accept has been revoked, since RFC 9901 defines no status check?
4. Which validity claims, such as iss, exp and cnf, do our credential profiles keep out of selective disclosure?
5. How long do we keep presented SD-JWTs, including in log files?

## Sources

- [S1] [Selective Disclosure for JSON Web Tokens](https://www.rfc-editor.org/rfc/rfc9901) (IETF, RFC 9901, Standards Track, November 2025)

## Evidence

- [BI-FIND-2026-0414](https://insights.bostonidentity.com/ref/BI-FIND-2026-0414): AP2 v0.2 authorizes an agent's purchase with two signed objects and a chain of hashes. The Checkout Mandate is bound by a checkout_hash claim to a Checkout JWT the merchant signed; the Payment Mandate is bound to the hash of that same JWT; both travel as SD-JWTs with a versioned type claim (mandate.payment.1, mandate.checkout.open.1) that verifiers must match exactly, and each verifier, merchant, credential provider, network and processor, has its own MUST list of what to check. The merchant's Checkout JWT must use a non-deterministic signature such as ECDSA rather than Ed25519, so a hash cannot be reversed from a table. Verifiers always receive closed mandates whatever mode produced them. (from "Google's AP2 v0.2 lets a person sign spending limits once and leave; the agent then signs each purchase itself")
- [BI-FIND-2026-0477](https://insights.bostonidentity.com/ref/BI-FIND-2026-0477): Verifiable Intent chains three signed SD-JWT layers: the card issuer binds the user's key for about a year (L1), the user signs mandates for the agent (L2: final values for an immediate purchase, or constraints plus the agent's key for up to thirty days in autonomous mode), and the agent signs a payment mandate for the network and a checkout mandate for the merchant, each valid about five minutes (L3a, L3b). L3 carries no confirmation key, and "The absence of `cnf` signals that no further delegation is permitted", so an agent cannot pass its authority to another agent. The only narrowing check is that the agent's final values satisfy the user's constraints, and a verifier checks only the constraints disclosed to it: with no allowlist disclosed, the allowlist check is skipped. Eight constraint types are defined (merchants, line items, payees, amount range, budget, recurrence, agent recurrence, reference); budget and the two recurrence limits need the payment network to keep a running count. (from "Verifiable Intent stops an agent from handing its authority on, and has no way to take that authority back")
- [BI-FIND-2026-0440](https://insights.bostonidentity.com/ref/BI-FIND-2026-0440): The consent mechanism in EMVCo's framework is not EMVCo's. Section 6 presents Verifiable Intent v0.1 as one realisation of the framework's three layers and tells reviewers the authoritative text lives in a FIDO Alliance repository: three chained SD-JWT verifiable credentials, the first signed by a payment credential provider (the named candidate is EMVCo's own Card DPC), the second by the consumer through a wallet with the agent's key inside it, the third by the agent as closed checkout and payment mandates bound by hash to a merchant-signed checkout. That is the open-mandate and closed-mandate structure of Google's AP2 v0.2, down to the rule that an unrecognised constraint type must fail. The framework never names AP2, Google, Mastercard or Visa, declares SD-JWT out of scope in section 1.2.2, and specifies SD-JWT VC as the format in 6.7. Two of the three groups writing agent-payment rules have settled on one credential shape; the Agentic Commerce Protocol's allowance-bound vault token is the one that has not. (from "EMVCo's agent-payment framework is not a spec; its consent proof is FIDO's Verifiable Intent, in AP2's shape")
- [BI-CALL-2026-0077](https://insights.bostonidentity.com/ref/BI-CALL-2026-0077): draft-ietf-oauth-sd-jwt-vc is published as an RFC by 31 October 2027. The IETF tracker lists it in IESG evaluation, and revision 22, posted on 27 October 2026, answers IESG ballot comments from Ketan Talaulikar; IANA has yet to review the new revision. Revision 22 also makes the Token Status List draft a normative reference. That draft is already in the RFC Editor's queue, still without an RFC number, so we expect the new dependency to cost little time. We are wrong if no RFC has been published by that date. (from "Bad display metadata no longer gets an SD-JWT VC rejected, and Token Status List becomes normative")
- [BI-FIND-2026-0447](https://insights.bostonidentity.com/ref/BI-FIND-2026-0447): The draft is moving. The OAuth working group chairs opened an official call for adoption on 28 September 2026, replies due 12 October. We counted the list on 6 October: fifteen replies, every one in support and none opposed, among them Arnar Birgisson of Google and names from the SD-JWT VC and OpenID4VP work. One reply, supporting 'as a starting point', listed attenuation first among the things the group would have to address. The author's repository shows an editorial pass merged on 21 September that the datatracker's -00 does not carry. A working-group revision would make this the only delegation format in our corpus with a standards home: AAuth, Agentic JWT and the agent name service all remain individual drafts. (from "Delegate SD-JWT lets a verifier trace hand-offs back to the issuer, but never checks each one narrowed authority")
- [BI-FIND-2025-0143](https://insights.bostonidentity.com/ref/BI-FIND-2025-0143): RFC 9901 never mentions agents or delegation. Its holder "may refer to the actual user, the supporting hardware and software in their possession, or both", and section 9.9 covers a credential moving past its holder: any entity holding an SD-JWT can forward it to any third party that does not enforce key binding, and may remove disclosures as it does. The RFC says this "could be desirable" in some cases and gives a supply chain where each party can choose to pass on less. The set of disclosures in a bare SD-JWT is not integrity protected, so the receiver has no way to tell what was dropped, and, as we read the format, nothing in it names who forwarded it. An agent handing on authority under the bare RFC therefore looks like whoever presents the token, which is why the delegation designs in our briefings add a chain of their own; our briefing on the Delegate SD-JWT draft found it lets the key-binding JWT itself be an SD-JWT that names the next holder's key. (from "The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifier")
- [BI-FIND-2025-0144](https://insights.bostonidentity.com/ref/BI-FIND-2025-0144): Key binding is optional, and the RFC gives the decision to the verifier: it "MUST decide whether Key Binding is required for a particular use case before verifying", and must not base that on whether a key-binding JWT arrived, because an attacker can strip one off. Without binding, the RFC says, the credential "can be replayed by anyone who gets access to it", a malicious verifier can present it to another verifier that does not require binding, and disclosures can be added or removed with the SD-JWT still valid. With binding, the key-binding JWT must carry iat, aud, nonce and an sd_hash over the issuer-signed JWT and the chosen disclosures, so the disclosure set is signed as well. For agents, we read this as making replay protection a setting each verifier has to get right. One agent draft in our briefings does not keep the binding as written: the SD-Card draft, now expired, lists the nonce as an optional value where RFC 9901 makes it required. (from "The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifier")
- [BI-FIND-2026-0444](https://insights.bostonidentity.com/ref/BI-FIND-2026-0444): The mechanism is one substitution. In RFC 9901 the holder proves possession by appending a signed Key Binding JWT; here that slot can hold a KB-SD-JWT, a selective-disclosure token in its own right whose 'delegate payload' carries the claims an issuer would have written and, for any link that delegates onward, a cnf key naming the next holder. A verifier splits the compact form, validates the issuer's SD-JWT, then validates each KB-SD-JWT with the previous link's cnf as the issuer key, requires exactly one disclosed delegate payload, and checks a hash of the predecessor: sd_hash pins the exact disclosures that were passed on, issuer_jwt_hash lets the delegate drop some. Any failed step aborts. One signature may pre-authorise several alternative delegations, of which the delegate reveals one. The worked example is an AI agent buying at a merchant without revealing which other merchants could have filled the order. (from "Delegate SD-JWT lets a verifier trace hand-offs back to the issuer, but never checks each one narrowed authority")
- [BI-FIND-2026-0632](https://insights.bostonidentity.com/ref/BI-FIND-2026-0632): Token Status List replaces per-token checks with one signed, compressed bit array per issuer. Each token (JWT, SD-JWT VC, CWT or ISO mdoc) carries an index and a list URI; a verifier fetches the list, checks its signature and reads 1 to 8 bits for VALID, INVALID, SUSPENDED or an application value. Because everyone downloads the same list, the issuer does not learn which token is being checked, the property the authors call herd privacy, unlike OAuth introspection or OCSP. Revocation reaches a verifier only when it refetches, which the issuer steers with the list's exp and ttl. (from "OAuth's revocation list for issuers is nearly an RFC, and agent drafts we read that need revocation don't use it")
- [BI-FIND-2025-0145](https://insights.bostonidentity.com/ref/BI-FIND-2025-0145): What RFC 9901 hands to the protocol or profile on top is, as we read it, most of what an agent deployment has to settle: how the holder's key pair is established (section 4.1.2), how the key-binding JWT's audience is represented and how its nonce is obtained (4.3), if and how consent to disclose is obtained (7.2), which validity claims a verifier must see, which application-specific profiles SHOULD list (9.7), and how verifiers get issuers' keys, which it does not define (9.8). It defines no status or revocation check for an issued SD-JWT; Token Status List appears only as an example of a revocation callback that tells the issuer little. In our briefings these gaps mostly carry forward: Verifiable Intent v0.1 and the Delegate SD-JWT draft define no revocation, and the SD-Card draft puts the registry that issues agent cards out of scope. Our briefing on SD-JWT VC found that it answers the issuer-key question with a well-known metadata document or an X.509 chain. (from "The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifier")
- [BI-FIND-2025-0230](https://insights.bostonidentity.com/ref/BI-FIND-2025-0230): In December 2025 the OpenID Foundation said implementers could start self-certifying against OpenID4VP 1.0, OpenID4VCI 1.0 and HAIP 1.0 on 26 February 2026, using free, open-source tests that run locally or on its servers. The two suites were at different stages. Presentation tests already covered both SD-JWT and mdoc credentials, but issuance tests covered SD-JWT only, with mdoc support "planned for release in the weeks ahead." Sending test logs to the Foundation for review and publication is optional. Accreditation of local bodies and labs was to follow in the second quarter of 2026. (from "The OpenID Foundation set 26 February 2026 for verifiable credential self-certification, with issuance tests still SD-JWT only")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/sd-jwt as the source. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
