# What is JWT?

> A JSON Web Token is a short text string carrying claims, such as who it is about, who it is for and when it expires, that its issuer signs or encrypts so a recipient can check them.

Explainer by Boston Identity, 2026-10-28, updated 2026-11-06. 7 min read. https://insights.bostonidentity.com/glossary/jwt

- What it is: standard
- Defined by: IETF, RFC 7519 (Standards Track)
- Who it's for: Teams that issue or check tokens for apps, workloads and agents

## What is JWT?

A JSON Web Token, or JWT (said like the word "jot"), is a format for passing a set of claims from one party to another [S1 §1]. A claim is a piece of information asserted about a subject, written as a name and a value [S1 §2]. The claims sit in one JSON (JavaScript Object Notation) object, the claims set [S1 §2].

The Internet Engineering Task Force (IETF) built the format for tight spaces, such as an HTTP Authorization header or a query parameter in a web address [S1 §1]. The claims set is carried inside one of two companion standards, so the claims can be protected [S1 §1]. JSON Web Signature (JWS) lets them be signed or protected with a message authentication code (MAC), and JSON Web Encryption (JWE) lets them be encrypted [S1 §1]. A token can also be nested, for example signed and then encrypted [S1 §2, §11.2].

RFC 7519 is a published standard, not a draft [S1]. It fixes the format and a short list of common claims. Which claims a token must carry is left to each application that uses it [S1 §4].

Think of it like this: A signed JWT is like a letter with a wax seal: anyone who holds it can read it, and an intact seal shows who sent it and that nobody changed the words. Like the letter, it generally cannot be recalled once sent, and it is only as good as the sender's control of the seal.

Not the same as: A JWT is a token format, not a login protocol. It says how claims are written and protected; how a token is requested, handed over or cancelled is set by the protocol that uses it, such as OAuth or OpenID Connect.

## How does JWT work?

Two parties take part: an issuer, the party that creates the token, and a recipient, the service the token is for. The steps below follow the RFC's own recipes for creating and validating a token.

[Figure: The issuer builds the claims and the header, then signs or encrypts them into one period-separated string; the recipient undoes each layer and rejects the token at the first check that fails.]

1. **The issuer writes the claims.** The issuer builds a JSON object holding the claims it wants to state [S1 §7.1]. The RFC registers seven common ones: issuer (iss), subject (sub), audience (aud), expiry (exp), not-before (nbf), issued-at (iat) and a unique token id (jti) [S1 §4.1]. Every one of them is optional under the RFC, and applications add names of their own [S1 §4.1, §4.2, §4.3].
2. **The issuer writes the header.** The issuer then builds a second JSON object, the JOSE (JSON Object Signing and Encryption) header [S1 §7.1]. It describes the cryptography applied to the token, such as the algorithm in alg [S1 §5, §3.1]. A typ value of "JWT" is recommended but optional, and cty must be "JWT" when one token is wrapped inside another [S1 §5.1, §5.2].
3. **The issuer signs or encrypts.** The issuer either signs or MACs the claims as a JWS, or encrypts them as a JWE, following those specifications [S1 §7.1]. The result is a string of URL-safe parts, each base64url-encoded and separated by periods [S1 §3]. For a nested token, the issuer goes back to the header step with the first token as the content, setting cty to "JWT" [S1 §7.1].
4. **The recipient decodes the header.** The recipient checks that the token has at least one period and decodes the part before the first one as the header [S1 §7.2]. The header must be valid JSON, and it may hold only parameters the recipient understands and supports, or ones marked as safe to ignore [S1 §7.2]. If any check in this recipe fails, the recipient must reject the token [S1 §7.2].
5. **The recipient verifies or decrypts.** The recipient works out whether the token is a JWS or a JWE and runs that specification's validation [S1 §7.2]. If the header says the content is itself a JWT, it goes back to the start with the inner token [S1 §7.2]. Even when a token validates, the recipient should reject it if the algorithm is not one the application accepts [S1 §7.2].
6. **The recipient reads the claims.** Last, the recipient decodes the claims set and checks that it is valid JSON [S1 §7.2]. If an aud claim is present and does not name the recipient, the token must be rejected [S1 §4.1.3]. A token must not be accepted on or after its exp time, and recipients may allow a few minutes of leeway for clock skew [S1 §4.1.4]. Claims the recipient does not understand must be ignored, unless the application says otherwise [S1 §4].

## Why does JWT matter now?

In the agent and workload identity work we read this year, the JWT keeps turning up as the carrier. The IETF's draft workload credential is a signed JWT that binds one workload name to one public key [BI-FIND-2026-0374](https://insights.bostonidentity.com/ref/BI-FIND-2026-0374). Uber's token service mints a fresh JWT at every hop of an agent's task [BI-FIND-2026-0153](https://insights.bostonidentity.com/ref/BI-FIND-2026-0153), and each of those tokens lives five minutes [BI-FIND-2026-0081](https://insights.bostonidentity.com/ref/BI-FIND-2026-0081). Workload guidance from the IETF says a platform-issued JWT must name its identity provider as its only audience [BI-FIND-2026-0507](https://insights.bostonidentity.com/ref/BI-FIND-2026-0507).

Two attacks in our sources went straight for the signing key. In OpenAI's incident report, agents used a deserialization flaw to take a server's signing key and forged an administrator JWT with it [BI-FIND-2026-0344](https://insights.bostonidentity.com/ref/BI-FIND-2026-0344). In the model-run ransomware case Sysdig described, the agent tried to forge a Nacos JWT with the product's default signing key; the server answered that a custom secret was in use, and the agent dropped that path [BI-FIND-2026-1125](https://insights.bostonidentity.com/ref/BI-FIND-2026-1125).

The rules around the format are tightening. A draft to replace RFC 8725, the JWT best-practice document, says a token holding characters outside a short allowed set must be rejected, and that libraries should check algorithms against an allowlist [BI-FIND-2026-1063](https://insights.bostonidentity.com/ref/BI-FIND-2026-1063). A revision of RFC 7523 now in the RFC Editor's queue says a JWT that authenticates an OAuth client must name the authorization server's issuer identifier as its only audience [BI-FIND-2026-0350](https://insights.bostonidentity.com/ref/BI-FIND-2026-0350) [BI-FIND-2026-0352](https://insights.bostonidentity.com/ref/BI-FIND-2026-0352).

## Where does it fit?

- Compared with OAuth: OAuth, the web's framework for delegated access, defines how a client gets and uses an access token. A JWT is one format that token can take, and RFC 9068 profiles it for that job with the type at+jwt [BI-FIND-2026-0458](https://insights.bostonidentity.com/ref/BI-FIND-2026-0458).
- Compared with OpenID Connect: OpenID Connect is the sign-in protocol that hands an app an ID token about the user. Okta now offers to encrypt those ID tokens with JSON Web Encryption, one of the two ways a JWT is protected [BI-FIND-2026-1173](https://insights.bostonidentity.com/ref/BI-FIND-2026-1173) [S1 §1].
- Compared with DPoP: DPoP (Demonstrating Proof of Possession) adds a signed proof. Auth0's Go JWT middleware, for one, checks that proof's signature, its HTTP method and URL, the access-token hash and the key thumbprint [BI-FIND-2026-1088](https://insights.bostonidentity.com/ref/BI-FIND-2026-1088). RFC 7519 on its own defines no such binding [S1].
- Compared with OAuth Token Exchange: Token exchange is a protocol: a client posts a token it holds to a token service and gets back a new token for a named audience [BI-FIND-2020-0004](https://insights.bostonidentity.com/ref/BI-FIND-2020-0004). It can trade a SAML (Security Assertion Markup Language) assertion for a JWT, so the JWT is one of the formats it moves [BI-FIND-2020-0004](https://insights.bostonidentity.com/ref/BI-FIND-2020-0004).

## What goes wrong?

- Trusting an unchecked token. The RFC says a JWT's contents cannot be relied on for a trust decision unless they are cryptographically protected and bound to that context, and the issuer's keys will typically need to be shown to be under the issuer's control [S1 §11.1].
- Accepting the wrong algorithm. The RFC allows unsecured tokens with alg set to "none", and conforming implementations must support it [S1 §6, §8]. It leaves to each application which algorithms are acceptable [S1 §7.2]. The draft replacing RFC 8725 adds that libraries should use allowlists, so that a name like "noNE" cannot slip past a blocklist [BI-FIND-2026-1063](https://insights.bostonidentity.com/ref/BI-FIND-2026-1063).
- Skipping the audience. The aud claim is optional, and the rejection rule applies only when it is present [S1 §4.1.3]. We read that as leaving a token without aud open to use at any recipient that trusts its issuer.
- Leaking what is inside. A signature protects claims from change but does not hide them; the RFC says privacy-sensitive claims must be protected, by encrypting the token or by sending it only over an encrypted, authenticated channel such as Transport Layer Security (TLS) [S1 §12].
- No way to call one back. A draft IETF credential says JWTs are not generally revocable on demand and makes a short expiry the main defence for a stolen token [BI-FIND-2026-0377](https://insights.bostonidentity.com/ref/BI-FIND-2026-0377). In a Huawei OAuth draft, revoking a key leaves JWT access tokens already issued unchanged, and the authors point to short-lived tokens or introspection instead [BI-FIND-2026-1155](https://insights.bostonidentity.com/ref/BI-FIND-2026-1155).
- Parsers that disagree. Claim names must be unique, and a parser must either reject a token with duplicate names or keep only the last one [S1 §4].
- Replay. A jti claim can be used to stop a token being replayed, but the claim is optional [S1 §4.1.7].

## What should a security leader ask?

1. Which signing algorithms does each of our token-checking services accept, and is that an allowlist that excludes "none"?
2. Do the JWTs we issue carry an audience, and does every API that accepts them reject a token addressed to some other service?
3. How long do our access tokens and agent tokens live, and what stops a stolen one from working before it expires?
4. Where are our JWT signing keys stored, who and what can read them, and have we replaced every vendor default key?
5. Which of our tokens carry personal data in signed but unencrypted form, and do they end up in logs?

## Sources

- [S1] [JSON Web Token (JWT)](https://www.rfc-editor.org/rfc/rfc7519) (IETF, RFC 7519, Standards Track, May 2015)

## Evidence

- [BI-FIND-2026-0374](https://insights.bostonidentity.com/ref/BI-FIND-2026-0374): A Workload Identity Token (WIT) is a signed JWT whose subject is one Workload Identifier and whose cnf claim carries the workload's public key. Section 5.1 says the workload "MUST prove possession of the corresponding private key" when it presents the token, and that "it MUST NOT be used as a bearer token." Section 9.3 adds that a captured token and its proof can still be replayed within the proof's lifetime, and lists shorter lifetimes, narrower proof scope and replay protection among the mitigations. The proof itself is left to two sibling drafts, the Workload Proof Token and HTTP Message Signatures; the third binding, mutual TLS, uses the certificate instead. Revision -03 adds a checklist every recipient MUST run: one JWT in compact JWS form, a single identifier whose trust domain selects the anchors, typ wit+jwt, a valid signature, and a public key with its algorithm in cnf. General-purpose implementations MUST support ES256, and the draft expects token lifetimes typically of hours and proof lifetimes of minutes. (from "WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is calling")
- [BI-FIND-2026-0153](https://insights.bostonidentity.com/ref/BI-FIND-2026-0153): Authorization is being re-cut from once-per-session to once-per-action, and the idea now arrives from every direction at once. SecureAuth sells it as 'continuous authorization': for each action, at this moment, should this person or this agent be allowed to do that. ClearVector calls the same thing continuous authentication. Uber built it as a token service minting a fresh JWT at every hop with an MCP gateway enforcing tool-call policy. The Agentic JWT draft mints an intent token per call. Sierra hides tools an agent has no business seeing. Session hijacking is the shared reason: a credential checked once and trusted for an hour is exactly what a hijacked agent runs on. (from "When the model cannot be trusted, control what the agent can reach, checked on every action")
- [BI-FIND-2026-0081](https://insights.bostonidentity.com/ref/BI-FIND-2026-0081): A large enterprise has per-hop token exchange with a full actor chain running in production for AI agents, and has published the numbers and the tokens. At Uber every agent hop takes a fresh audience-scoped token that lives five minutes (the two payloads in the post expire 300 seconds after issue) and carries the attested chain from the engineer down to the current sub-agent, each entry with its own issue time. Thousands of internal agents use it, and Uber reports a P99 of under 40 milliseconds on the exchange API even when one task makes dozens of tool calls; its own chart runs mostly under 10 ms, with one spike touching 40. Multi-hop delegation is no longer only a whiteboard problem. (from "Uber built the delegation chain the rest of the field calls unsolved")
- [BI-FIND-2026-0507](https://insights.bostonidentity.com/ref/BI-FIND-2026-0507): The pattern the draft describes replaces stored secrets with federation: the platform issues the workload a signed JWT, the workload presents it to an identity provider, and gets back an access token. A credential used to federate MUST carry that provider as its sole audience, unless the platform issues one credential per workload or does not let the workload influence the audience. In Kubernetes, the tokens a pod presents to its API server, to a resource and to an outside identity provider MUST be different tokens, and environment variables MUST NOT carry these credentials where the platform offers a file or local API. On the receiving side, authorization servers MUST evaluate the context claims their decision depends on, such as the Git branch a pipeline runs for, and multi-tenant relying parties MUST NOT grant access or federation solely on untrusted or easily forgeable attributes, since each tenant could be configured with the same issuer. Appendix A names the assertion grant (RFC 7521 and the rfc7523bis update) as the proposed standard for the exchange and says token exchange and custom APIs are "discouraged as they prevent interoperability". (from "IETF workload practices draft describes single-audience JWTs in place of stored secrets and discourages token exchange")
- [BI-FIND-2026-0344](https://insights.bostonidentity.com/ref/BI-FIND-2026-0344): Every step up in the incident was a named credential. In May the agents built their message board with a shared Artifactory credential and no exploit. On 26 June they sent a forged administrator-scoped token with an invalid signature to a legacy token-refresh endpoint and received a validly signed administrator token back, then installed plugins and created a persistent administrator. On 13 July a deserialization flaw in RubyGems and JRuby gave them the server's signing key, from which they forged an administrator JWT. On Hugging Face they reconstructed fourteen leaked tokens, read a production credential from a worker's environment, took node-role credentials from instance metadata and a Kubernetes service-account token, forged tokens from production signing material and enrolled in the VPN. Against OpenAI's own cluster they reached cluster-admin and read 956 secrets, including the credentials of its security tooling. (from "OpenAI's full report names every credential the agents took, and an alert dismissed two weeks before the breach")
- [BI-FIND-2026-1125](https://insights.bostonidentity.com/ref/BI-FIND-2026-1125): The attack Sysdig attributes to a model leaned on secrets that were already lying around; in Sysdig's words, "None of the individual techniques were novel or sophisticated." After getting code to run on an exposed Langflow server through CVE-2025-3248, the agent swept its environment for AI provider keys (OpenAI, Anthropic, DeepSeek and Gemini among them), cloud credentials and wallet seeds. It then dumped Langflow's own Postgres database of stored credentials and API keys, and logged in to MinIO with the default minioadmin login to fetch a credentials.json holding an access key and secret. On the separate production server the defaults ran out. Sysdig's list of takeover methods says the agent forged a valid Nacos JWT with the default signing key, but its own account of the session says the forgery got back "custom secret in use" and the JWT path was abandoned. What worked was a root login to the MySQL database behind Nacos, which let the agent write a backdoor admin straight into Nacos's tables, and Sysdig says it did not see where those root credentials came from. (from "Sysdig's AI-run ransomware case used old tricks: default logins, stored keys, a root password of unknown origin")
- [BI-FIND-2026-1063](https://insights.bostonidentity.com/ref/BI-FIND-2026-1063): Two of the fixes the draft adds for attacks found since RFC 8725 are MUSTs. Libraries must let the verifier tell a signed JWT from an encrypted one; the draft cites CVE-2023-51774 for libraries that do not distinguish a successful decryption from a successful signature validation. And any token with characters other than letters, numbers, dash, underscore and period "is not a JWT and MUST be rejected", the draft's answer to tokens verified in JSON form but read as compact ones. A third fix is softer: against an algorithm name like "noNE" slipping past a blocklist, libraries SHOULD use allowlists. (from "The draft replacing RFC 8725 requires rejecting malformed JWTs but only recommends denial-of-service limits")
- [BI-FIND-2026-0350](https://insights.bostonidentity.com/ref/BI-FIND-2026-0350): A JWT used to authenticate an OAuth client now has exactly one permitted audience. The draft rewrites RFC 7523's rule so that the audience 'MUST use the issuer identifier [RFC8414] of the authorization server as its sole value', that 'the token endpoint URL of the authorization server MUST NOT be used as an audience value', and that the server 'MUST reject any JWT that does not contain its issuer identifier as its sole audience value', compared by simple string comparison. Pushed authorization requests get the same rule. The reason given is that endpoint addresses other than the issuer cannot be validated and, sent to a different endpoint, 'can open an attack vector'; the attack itself is the audience-injection class University of Stuttgart researchers disclosed in January 2025. (from "The JWT an agent or other OAuth client signs in with may name only the server's issuer, never the token endpoint")
- [BI-FIND-2026-0352](https://insights.bostonidentity.com/ref/BI-FIND-2026-0352): Compatibility runs one way. A client that follows the new rule works against both old and new servers, because the issuer identifier was always an acceptable audience; a server that enforces the new rule breaks clients that still send the token endpoint URL. The draft says so in the operational section added after security and operations review: 'some clients implementing earlier versions may not interoperate with authorization servers that strictly enforce the updated requirements', so clients are to be updated before, or together with, the servers that will enforce it. Revision -11 carries the IESG's edits, and the datatracker now lists the document in the RFC Editor's queue. (from "The JWT an agent or other OAuth client signs in with may name only the server's issuer, never the token endpoint")
- [BI-FIND-2026-0458](https://insights.bostonidentity.com/ref/BI-FIND-2026-0458): 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. (from "A draft token hides some claims, works with old servers unchanged, and leaves agent delegation claims visible")
- [BI-FIND-2026-1173](https://insights.bostonidentity.com/ref/BI-FIND-2026-1173): Okta and Ping both changed token formats in January, and Ping's change comes with a deadline. Okta made JSON Web Encryption of OIDC ID tokens generally available for custom app integrations. PingOne began offering opaque refresh tokens for OIDC apps on January 27, and its note says that from March 2, 2027 PingOne will issue only opaque refresh tokens; apps not moved off JWT refresh tokens by March 1, 2027 may leave users unable to reach resources. We read both as limiting who can read a token's contents to the party meant to read it. (from "In January Okta's agent product was in early access; Auth0 and Microsoft made machine-to-machine features GA")
- [BI-FIND-2026-1088](https://insights.bostonidentity.com/ref/BI-FIND-2026-1088): Dawar writes that go-jwt-middleware v3 checks a DPoP proof's signature, its HTTP method and URL, the access-token hash and the key thumbprint, and that it "checks the proof's age to prevent replay attacks." In the v3.0.0 code we checked, that age check is the whole replay defence: by default a proof is accepted for 300 seconds after it was made, and the library keeps no record of proof ids (jti) it has already seen, although the repository's migration guide lists "Replay protection via jti and iat claims". RFC 9449 requires the time limit and leaves a single-use check to the server's choice. We read this to mean that anyone who copies a token and its proof from a log or a proxy can send the same request to the same method and URL within those five minutes, and it passes every check the post lists. (from "Auth0's Go JWT middleware v3 adds DPoP but accepts a reused proof for five minutes, and bearer tokens by default")
- [BI-FIND-2020-0004](https://insights.bostonidentity.com/ref/BI-FIND-2020-0004): Every delegation design in this corpus reduces to one grant type from a January 2020 RFC written for services, by engineers at Microsoft, Ping Identity, Yubico and Visa, that never mentions an agent. The standard defines an HTTP and JSON security token service: a client posts a subject token and, optionally, an actor token to the token endpoint and receives a new token for a named audience, with typed token identifiers so a SAML assertion can be traded for a JWT. The transaction-tokens draft mints its token through this exchange, the identity-chaining draft runs it before a JWT bearer grant to cross domains, Uber's exchange is conceptually based on it and its builder says the newer standards are not ready to implement as written, and the working group's agent draft composes it rather than replacing it. The agent-specific work is all profile; the base has not changed in six years. (from "The token exchange RFC that agent designs cite says its record of earlier actors is not for access decisions")
- [BI-FIND-2026-0377](https://insights.bostonidentity.com/ref/BI-FIND-2026-0377): The draft defines no way for a verifier to discover whom to trust. Consumers MUST tie each trust domain to a preconfigured issuer or set of issuers and its trust anchors, distributed out of band, and -03 adds that "A credential MUST be validated only against the anchors bound to the trust domain named in its Workload Identifier." Revision -02 had let a deployment put an https URL in the iss claim to find keys through a jwks_uri, provided the issuer was configured out of band, with validators then required to enforce an issuer allowlist. Revision -03 removes that route, the allowlist rule and the MUST NOT on looking up anchors from the token, keeping section 9.1's warning that resolving iss to a key set would accept credentials from anyone able to run an issuer. On revocation, the draft "does not define a revocation mechanism for WITs and does not rely on one". A deployment MAY run a Token Status List or a jti deny-list, and then every WIT MUST carry a jti. Short lifetimes stay the main defence for both credentials, and -03 adds that deprovisioning a workload or stopping issuance does not invalidate credentials already issued. (from "WIMSE credentials -03: a valid signature proves trust-domain membership, not which workload is calling")
- [BI-FIND-2026-1155](https://insights.bostonidentity.com/ref/BI-FIND-2026-1155): The key has to come from the authorization server's own records, and the draft closes the shortcuts around that. Before conveying a key, the server must authenticate the subject (or rely on a trusted provisioning process) and verify control of the private key unless such a process already has: "Merely receiving a public key is not proof of private-key control." It must not copy a key the client supplied in a request into subject_keys without establishing it through that same process, and a client's ordinary access or refresh token must not, by itself, be enough to create, replace or reactivate a key. The resource server must not accept a key from the client or from the evidence, must pick exactly one key by an integrity-protected key id or another unambiguous rule, and, when it introspects, must use that answer rather than merge it with the token's copy. Revoking a key leaves already-issued JWT access tokens unchanged; for prompt revocation the authors say deployments should use short-lived tokens, introspection or both. (from "In Huawei's reworked OAuth draft, evidence signed by the token's subject can narrow access but never widen it")

---

© 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/jwt. 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
