WIT
Workload Identity Token
By Boston Identity · Explained · 6 min read
The short answer
A Workload Identity Token is a signed token that names one piece of software and carries its public key. The software must prove it holds the matching private key each time, so a stolen copy alone is not enough.
- What it is
- A standard
- Defined by
- IETF WIMSE WG draft (-03), not yet a standard
- Who it's for
- Teams securing service-to-service and agent calls
- In our briefings
- 192by month, last 12
What is WIT?
A Workload Identity Token, or WIT, is the token form of a workload's identity in WIMSE (Workload Identity in Multi System Environments), an IETF working group. The WIMSE architecture covers authentication between software workloads, from basic runtime environments to multi-cloud, multi-tenant deployments.S1 The WIT is a JSON Web Token (JWT) signed by an Identity Server, and it binds a public key to the workload's identifier.§5.1 Its companion, the Workload Identity Certificate (WIC), does the same job as an X.509 certificate.§1
The problem it solves is calls between workloads at the application layer, where end-to-end transport security such as TLS (Transport Layer Security) is not always possible.§5 The WIT is for those application-layer calls; the certificate is aimed mainly at mutual TLS.§1 The token is never enough on its own. A workload presenting it must prove possession of the private key, and the draft says it must not be used as a bearer token, one that works for whoever holds it.§5.1
The document is an Internet-Draft at revision -03, published on 30 October 2026, so it is a draft, not yet a standard.S1 The proof itself is defined in separate documents.§1
A WIT is not a bearer JWT, which works for whoever holds it. The draft says a WIT must not be used as a bearer token, though during a migration a workload may accept both kinds from callers.§5.1,§5.4
How does WIT work?
Two workloads take part, a caller (Workload A) and a callee (Workload B), along with the Identity Server that issues the token. They belong to the same trust domain, the security boundary that names who may issue identities.§1.1
Figure steps
- The operator sets up trust anchors
- The Identity Server issues the WIT
- Workload A sends the WIT with a proof
- Workload B validates the token
- Workload B checks the proof is meant for it
- Workload B authorizes the caller and responds
Workload B trusts the WIT only after checking it against trust anchors it already holds and confirming that the caller's proof was addressed to it.
Drawn from draft-ietf-wimse-workload-creds-03, sections 1.1, 3 and 5
Swipe for the full figure →
The operator sets up trust anchors
Before any call, the callee must know whom to trust. Each trust domain is tied to a preconfigured issuer, or set of issuers, and its trust anchors, such as a set of public keys.§3 That mapping must arrive through a secure out-of-band channel, because the draft defines no way to discover it.§3
The Identity Server issues the WIT
The Identity Server signs a JWT for Workload A. Its sub claim holds exactly one Workload Identifier, a URI that includes the trust domain, and its cnf claim holds the workload's public key with the algorithm the proof must use.§5.1 The token must have a limited lifetime set by its exp claim, and the draft says it should be short-lived, typically on the order of hours.§9.3.1
Workload A sends the WIT with a proof
Workload A calls Workload B over a server-authenticated TLS connection.§1.1, §9.6 It puts the WIT in an HTTP header named Workload-Identity-Token, and it adds a proof of possession signed with its private key.§5.1.1, §1 The proof is either a Workload Proof Token or an HTTP Message Signature, each defined in its own draft.§1
Workload B validates the token
Workload B must check that the token is a single, well-formed JWT and that sub holds one Workload Identifier.§5.1.3 It uses that identifier's trust domain to pick the trust anchors. It must also check that the typ header is wit+jwt, that the signature is valid under those anchors, and that cnf holds a key with an algorithm.§5.1.3 It ignores claims it does not understand.
Why does WIT matter now?
Implementers already run the certificate half of WIMSE in production, while none of those listed reports the token half at production yet.1 In -03, the draft's own implementation list puts the token at beta at SPIFFE, alpha at Defakto Security and research at Teleport, while Defakto and Teleport list the certificate at production.1 SPIFFE has added WIT to its core specification as a third credential format, but its separate WIT-SVID profile is still marked Incubating, where breaking changes may follow implementation experience.2 We forecast that by 31 December 2027 at least one of SPIFFE, Defakto Security or Teleport will document the token at production or general availability.3
Two agent drafts already build on the token. An agent authentication draft names WIT among the credential formats an agent may hold, and calls static API keys an antipattern.4 An agent card draft says a card's subject and key should match the agent's WIT.5 Separately, an OAuth working-group draft lets a workload authenticate to an authorization server with a WIT-SVID, which requires proof of a private key, unlike a replayable JWT-SVID.6,7 The WIT itself still carries no delegation claim, and the word agent does not appear in the -03 text.8
Where does it fit?
| Compared with | How WIT differs |
|---|---|
| WIMSE Workload Proof Token | The Workload Proof Token is one of the two proofs that can travel with a WIT; the other is an HTTP Message Signature.§1 The WIT says who the workload is, while the proof shows the caller holds the key and is typically valid for minutes.§9.3.2 |
| JWT-SVID | A JWT-SVID is SPIFFE's signed identity token, and an attacker who obtains one can replay it until it expires at the server it names.7 A WIT carries a public key and must not be used as a bearer token: the workload must prove it holds the matching private key.§5.1 |
| SPIFFE | Kasselman, a WIMSE participant, says WIMSE borrows from SPIFFE and makes it a little more general.9 A SPIFFE deployment uses spiffe URIs as the identifier in a WIT, while the wimse URI scheme serves deployments without an environment-specific scheme.§5.1 |
What goes wrong?
- A valid signature proves the caller belongs to the trust domain, not which workload it is. The callee still has to compare the identifier it expects, and proof of possession does not close that gap.§9.2
- A captured WIT and its proof can be replayed within the proof's lifetime. The draft lists short proof lifetimes, a narrow proof scope and replay protection as mitigations.§9.3
- There is no revocation by default. The draft relies on short lifetimes, and deprovisioning a workload does not invalidate the tokens it already holds.S1 A deployment may run a Token Status List or a jti deny-list, and then every WIT must carry a jti.§9.3.1
- Resolving the iss claim to fetch keys would accept credentials from anyone able to run an issuer.§9.1
- The draft requires no particular key protection, so verifiers must not infer hardware-backed keys from the WIT alone.§9.5
- Intermediaries see the token and proof in the clear and can try to replay them.§9.6
- Migration has a trap. A workload may accept both bearer tokens and WITs for a while, but it picks which token to prefer before the caller is authenticated, so that choice has to be checked again afterwards.§5.4
What should a security leader ask?
- 01
When one of our workloads is deprovisioned, how long can its existing tokens still be accepted, and who chose that lifetime?
- 02
Do our services compare the caller's full Workload Identifier against what they expect, or do they stop at a valid signature from our trust domain?
- 03
Which of our vendors' WIT support is production, and which is beta, alpha or research?
- 04
Where do our agents carry delegation and caller context, given that the WIT carries neither?
- 05
How do our verifiers receive each trust domain's anchors, and could any of them fetch keys from a URL in the token?
Sources
- S1
WIMSE Workload Credentials (draft-ietf-wimse-workload-creds-03)IETF WIMSE Working Group · Internet-Draft -03, 30 October 2026
›Evidence: 9 claims from 6 briefings
- 1BI-FIND-2026-0375 · WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is callingThe Workload Identity Certificate (WIC) is an ordinary X.509 certificate with the Workload Identifier in exactly one URI SubjectAltName. The implementation section calls it "fully compatible with the X509-SVID and widely used" for SPIFFE, whose deployments use spiffe URIs as the identifier. Revision -03 also says plainly that mutual TLS uses this certificate, not the WIT, even when a WIT travels over a TLS channel. The token is the new piece. Its self-reported status is beta at SPIFFE, alpha at Defakto Security and research at Teleport, while Defakto and Teleport list the certificate at production. A fourth entry, new in -03, is wimsey, an independent project that lists both forms at alpha. We read this as the IETF standardizing a certificate the industry already issues and adding a token that no listed implementer yet reports at production.
- 2BI-FIND-2026-2085 · SPIFFE trust domain names have no registrar, and the core spec warns that roles in a credential can go staleThe core spec lists WIT as a third SVID format next to X.509 and JWT, and its text does not say what WIT is or that it is less settled than the other two. A commit on 1 July 2026 added WIT to that list, and the core file was marked Stable the same day. The separate WIT-SVID file in the repository, which we checked, is marked Incubating, while the X.509 and JWT files are marked Stable. The project's stability document says that at Incubating, breaking changes are avoided but "may be made in response to real-world implementation experience", and that features relying on such a spec are "typically gated behind a feature flag". The WIT-SVID file describes itself as a profile of the Workload Identity Token from the IETF's WIMSE working group.
- 3BI-CALL-2026-0063 · WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is callingBy 31 December 2027 at least one of three implementers the draft lists (SPIFFE, under the CNCF; Defakto Security; Teleport) documents the JWT form of the credential, the Workload Identity Token, at production or generally-available status in its own release notes or documentation, naming this draft or its RFC number. Revision -03 still lists the token at beta, alpha and research at those three, while the certificate is widely used or in production, so the token is the piece left to ship. Defakto Security also employs one of the draft's authors.
- 4BI-FIND-2026-0179 · Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reachThe draft's sharpest rule reads "The Large Language Model MUST NOT have access to an agent's credentials or to credentials that may be needed to access tools and services", which it says keeps prompt injection from getting the model to use, expose or disclose them. Around it sit the identity rules. An agent MUST be assigned exactly one WIMSE identifier and MUST hold credentials cryptographically bound to it; those credentials SHOULD be short-lived and MUST carry an explicit expiry, and static API keys are "an antipattern for agent identity". The draft names WIT, X.509 WIC and SPIFFE SVIDs as formats and leaves the choice to the deployment. Our briefing on Sierra's cloud agent platform found a similar boundary built in: a proxy approves each call and swaps in the credential, so the agent never holds it.
- 5BI-FIND-2026-1959 · Cisco's SD-Card rewrite pins MCP tools to what a registry signed; its security sections still read TODOThe draft keeps the card's issuer apart from the workload identity issuer, and for delegation it uses the card itself. With WIMSE, a card's sub and cnf SHOULD match the agent's Workload Identity Token (WIT). In the authors' words a WIT "says which workload a request comes from", while the card says what that workload offers and is attested by a registry that may sit in a different trust domain, and a verifier MUST NOT treat either issuer as an issuer of the other's claims. It defines no delegation claim of its own: an agent MAY present its card as the actor token in OAuth token exchange (RFC 8693), the issued token names it in an act claim, and the authorization server MAY narrow the scope to capabilities the target agent's card discloses.
- 6BI-FIND-2026-0316 · OAuth's SPIFFE draft opens metadata URLs to all three credentials and says what a self-served key cannot proveAn OAuth working-group draft lets a workload authenticate to an authorization server with the SPIFFE identity its platform issued, instead of a stored client secret. Revision -03 keeps three forms: a signed JWT-SVID sent as a client assertion, an X509-SVID presented over mutual TLS, and a WIT-SVID sent under attestation-based client authentication. The certificate form is checked only against the anchors of its own trust domain and never the system trust store, because, the draft says, any certificate authority in that store could otherwise issue a trusted X509-SVID for any SPIFFE ID. A server must support at least one and must advertise the ones it accepts in its metadata, under three method names the draft asks IANA to register: spiffe_jwt, spiffe_x509 and spiffe_wit. One implementation is listed, a Keycloak preview whose stated coverage is the JWT-SVID form with a trust bundle endpoint.
- 7BI-FIND-2026-2024 · OAuth's SPIFFE draft opens metadata URLs to all three credentials and says what a self-served key cannot proveThe new replay section says an attacker that obtains a JWT-SVID can use it for client authentication until it expires, though only at the server named in its audience. SPIFFE permits but does not require a jti claim, and SPIFFE implementations may hand a workload the same cached JWT-SVID for several requests, so authorization servers cannot in general count on the jti replay check that RFC 7523 describes. Deployments that need replay resistance should use JWT-SVIDs with a unique jti that are never reused, or switch to the X509-SVID or WIT-SVID forms, which require proof of possession of a private key.
- 8BI-FIND-2026-0376 · WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is callingThe -03 text defines no delegation claim, no actor chain and no attestation evidence for either credential. Closed environments MAY add claims to the WIT freely, but elsewhere any new claim MUST be registered with IANA before use. Revision -03 draws the same line for keys: the draft requires no particular key protection, and "Verifiers must not infer such stronger properties from the WIT or WIC alone", meaning hardware-backed keys or binding to a measured runtime. Context belongs to another token. Section 5.4 says a WIT can coexist with a transaction token, and its example has the WIT's identity decide which API calls a workload may make and the context token which resources. The word agent does not appear in the -03 text. For agent designs that use the WIT as an agent's workload credential, we read this as meaning delegation and caller context have to travel separately.
- 9BI-FIND-2026-2133 · Two WIMSE participants say AI agents are workloads, and existing standards already cover most of the jobKasselman says WIMSE borrows from SPIFFE and makes it a little more abstract and a little more general. The SPIFFE idea, as he tells it, was to get rid of "secret zero": identify the workload at runtime, check that it runs in the right environment, and only then provision its credential, so no bootstrap secret is needed. The work flows both ways, he says: WIMSE's workload identity token "has now been adopted back into" SPIFFE. Rosomakho says the WIMSE identifier recently passed through the working group, and that a SPIFFE identifier fits within it while other schemes can too, such as ones hyperscalers might want instead. Kasselman's evidence that the model scales is second-hand: he recalls someone from Uber saying, he thinks at a public conference, that Uber was "issuing around a billion credentials a day".
›In our briefings
The claim each of the newest briefings on WIT would have you leave with.
Hulea's agent-mandate draft keeps evidence that a delegator may delegate apart from evidence that the delegator granted a mandate. An Authority Source attests which powers the delegator may delegate, and a Mandate Issuer signs the grant, bound to the agent's key; in the draft's words, "a Mandate Assertion alone never establishes the delegator's authority." No Authority Assertion is needed when the delegator is the principal, and -01 accepts that basis only for identifier schemes the Validation Profile admits. The protocol cannot verify the link between delegator and Mandate Issuer inside the delegator's organization; the draft leaves that link to the Mandate Issuer, under the trust framework's rules. Its security considerations add a gap: when another assertion establishes authority, an agent can simply omit a withdrawn Authority Assertion, the Validator then records only authority_not_presented, and "RP policy SHOULD examine diagnostics with the codes revoked, suspended, and authority_not_presented before allowing an action."
An agent-mandate draft checks who may delegate, and now intends to carry its chain in Asor's formatBI-FIND-2026-1576
Since -01 the draft has reversed which actors in a delegation chain count when a service admits an agent's call. Section 5's seven steps, for a Serving Party "that admits calls from agents only if they are registered and assured", now check the Principal and the current actor, the outermost act claim; -01 checked every Actor in the chain. Citing Section 4.1 of RFC 8693, the draft says prior actors "are informational only and MUST NOT be used to decide admission", and our reading of the RFC's own text agrees. At admission the chain is "the token issuer's asserted history", in the authors' words, neither complete nor evidence that any hop held or narrowed authority. Earlier actors stay in the record as audit history for Linked Verification to check. Showing that authority narrowed at each hop "remains an open issue" of the companion SCITT draft. The change log says the posted -01 had been uploaded from an earlier commit that lacked this change.
An agent-evidence draft no longer checks every actor in a delegation chain before a service admits a callBI-FIND-2026-1390
Rosomakho and Kasselman both treat an AI agent as a workload, so the identity standards built for workloads apply to it. Kasselman describes an agent as an algorithm in software "that runs on a processor somewhere", and says testing that idea was part of the thinking behind the AIMS draft the WIMSE group adopted, which "will not introduce any new work". The existing standards get you, in his words, "I'd say 80 90 95%" of the way. When people tell him the standards do not work, he says, they usually mean their vendor's product is not flexible enough to configure the fine-grained rules they need, or believe they need. The gaps Rosomakho names are particular flows, such as a sequence of agents passing context along, and he says they are no reason to start over. The host said he did not know that everybody accepts agents as workloads.
Two WIMSE participants say AI agents are workloads, and existing standards already cover most of the jobBI-FIND-2026-2132
›Cited in 19 briefings
Every briefing that discusses this term, newest first.
- An agent-mandate draft checks who may delegate, and now intends to carry its chain in Asor's format
Draft · · IETF Datatracker · Claudiu Hulea, Braincap SRL · Oct 24, 2026
- An agent-evidence draft no longer checks every actor in a delegation chain before a service admits a call
Draft · · IETF Datatracker · Shameer Thaha et al. · Oct 19, 2026
- Two WIMSE participants say AI agents are workloads, and existing standards already cover most of the job
Video · · Identerati Office Hours · Nov 6, 2026
- SPIFFE trust domain names have no registrar, and the core spec warns that roles in a credential can go stale
Article · · SPIFFE project (GitHub) · SPIFFE project · Jul 1, 2026
- OAuth's SPIFFE draft opens metadata URLs to all three credentials and says what a self-served key cannot prove
Draft · · IETF Datatracker · Arndt Schwenkschuster et al. · Jun 15, 2026
- Cisco's SD-Card rewrite pins MCP tools to what a registry signed; its security sections still read TODO
Draft · · IETF Datatracker · Suhas Nandakumar et al. · Nov 1, 2026
- A 'human approved' field is not evidence, says Schrock's draft; only a bound, signed approval counts
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Jul 3, 2026
- WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is calling
Draft · · IETF Datatracker · Brian Campbell et al. · Jul 2, 2026
- Agent URI draft -04 ties an agent's name to a key document and bars OAuth servers from trusting the name alone
Draft · · IETF Datatracker · Yaswanth Narvaneni et al. · Oct 20, 2026
- The IETF's stronger workload proof is built for the proxy in the middle, and it still leaves replay protection optional
Draft · · IETF Datatracker · Joseph A. Salowey et al. · Oct 20, 2026
- Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reach
Draft · · IETF Internet-Draft (individual submission, since adopted by the WIMSE working group) · Pieter Kasselman et al. · Jul 6, 2026
- Folding every hop's signature into one would stop anyone stripping a signed hop out of an agent's request
Draft · · IETF Datatracker · Tirumaleswar Reddy.K et al. · Sep 28, 2026
- The IETF's simplest workload proof proves the key and leaves the request unprotected, and the draft says so itself
Draft · · IETF Datatracker · Brian Campbell et al. · Aug 27, 2026
- The IETF's workload-identity architecture gives AI agents one subsection, calling an agent a delegated workload
Draft · · IETF Datatracker · Joseph Salowey et al. · Jul 6, 2026
- A June draft moves workload attestation out of TLS into HTTP headers so it survives a TLS-ending proxy
Draft · · IETF Datatracker · Tirumaleswar Reddy.K et al. · Jun 7, 2026
- Agent payment protocols prove an agent may pay; a Toronto paper adds a step to check the job was done
Paper · · arXiv · Adrian de Valois-Franklin et al. · Jun 7, 2026
- An Alibaba and Cisco draft signs what the user was shown, then enforces a policy the agent wrote
Draft · · IETF Datatracker · Dapeng Liu et al. · Feb 28, 2026
- A draft binds workload credentials to their host via a local agent, and leaves that agent's compromise undecided
Draft · · IETF Datatracker · Ram Krishnan et al. · Feb 26, 2026
- WIMSE leaves its application proof choice open, with different integrity coverage
Draft · · IETF Datatracker · Brian Campbell et al. · Oct 16, 2025
Related terms
What appears alongside WIT in the same briefings.
Was this explainer useful?
Share this explainer
Pass it on. Every claim has a permanent number, so what you share today still points at the same words next year.