Skip to content
IAM Insights

Explainers · Glossary

Proof of possession

PoP · proof-of-possession · PoP JWT

By Boston Identity · Explained · 6 min read

The short answer

Proof of possession ties a token to a key its holder keeps. To use the token, the holder must also prove it has that key, so a recipient can turn away someone who copied the token but lacks the key.

What it is
A design pattern
Defined by
IETF, RFC 7800 (for JWTs)
Who it's for
Teams issuing or accepting tokens for agents and workloads
In our briefings
50By month of the source: 2025-12: 0, 2026-01: 0, 2026-02: 3, 2026-03: 2, 2026-04: 1, 2026-05: 1, 2026-06: 3, 2026-07: 7, 2026-08: 3, 2026-09: 8, 2026-10: 4, 2026-11: 11by month, last 12
01

What is Proof of possession?

A bearer credential works for anyone who holds its bytes.1 Proof of possession (PoP) adds a second test: the token names a key, and whoever presents the token has to show it holds that key. RFC 7800, a standards-track document from the Internet Engineering Task Force (IETF) published in April 2016, defines this for a JSON Web Token (JWT), a signed set of claims. It says how a JWT declares that its presenter holds a particular key, and how the recipient can confirm that cryptographically.§1 The spec notes that this is also called being a holder-of-key.§1

The declaration lives in one claim, "cnf" (confirmation). By including it, the issuer of the token states that the presenter holds a particular key and that the recipient can check this.§3 The claim can carry the public key itself, an encrypted symmetric key, a key ID, or a URL for a set of keys.§3.2–3.5 It must represent only one proof-of-possession key.§3.1

RFC 7800 leaves the proof itself to other protocols. It says possession is typically shown by signing a value the recipient chose, a nonce or challenge, but it does not say how that value or the signature travels.§3.6

Not the same as DPoP, a proof-of-possession mechanism for OAuth tokens.2 RFC 7800 only says how a token names its key and leaves the proof to each protocol.§3.6

02

How does Proof of possession work?

RFC 7800 names three parties: the issuer, which creates the JWT and binds the key to it; the presenter, which proves it holds the key; and the recipient, which receives the JWT.§2 We follow the spec's example with an asymmetric key, a private key paired with a public one.§1

Figure steps
  1. The presenter makes a key pair
  2. The issuer binds the key into the token
  3. The issuer sends the token to the presenter
  4. The presenter shows the token and proves the key
  5. The recipient checks both
Fig. 1

The issuer writes the presenter's public key into the token, so the recipient can ask the presenter to prove it holds the matching private key.

Drawn from RFC 7800, section 1, Figure 2

Swipe for the full figure →

  1. The presenter makes a key pair

    The presenter generates a public and private key pair and sends the public key to the issuer.§1 In the spec's symmetric example, either side generates a shared secret key instead and sends it privately to the other.§1

  2. The issuer binds the key into the token

    The issuer creates a JWT whose cnf claim holds the public key, or an identifier for it.§1 It signs the whole JWT, which protects it against changes.§1 At least one of the "sub" (subject) and "iss" (issuer) claims must be present, and the presenter is normally the one they name.§3

  3. The issuer sends the token to the presenter

    The issuer sends the signed JWT to the presenter.§1 The presenter now holds the token and the private key that matches the key named inside it.§1

  4. The presenter shows the token and proves the key

    When the presenter hands the JWT to a recipient, it also has to demonstrate possession of the private key.§1 The spec gives two examples: use the key in a Transport Layer Security (TLS) exchange with the recipient, or sign a nonce, a one-time value.§1 Typically the recipient chooses the value to be signed.§3.6

  5. The recipient checks both

    The recipient first verifies the digital signature on the JWT.§1 It then takes the public key from the cnf claim and uses it to check the TLS exchange or the signed nonce.§1 In the spec's words, this lets the recipient verify "that it is interacting with the genuine presenter".§1

03

Why does Proof of possession matter now?

The agent-credential drafts we read this year split on this point, and several land on bearer tokens. One agent grant draft makes its assertion a bearer credential and lists proof of possession as an open issue.3 Another agent-credential draft defines no proof-of-possession binding, so its credentials work as bearer tokens valid for up to 24 hours.4 A paper on one agent runtime says it also uses bearer tokens with no proof of possession.5

An IETF draft on workload practices is firmer. Its version 08 says that where proof of possession is not supported, deployments must apply compensating controls such as short-lived tokens, narrow audiences or network controls.6 One agent payment protocol puts a shopping agent's public key in a cnf claim inside a mandate the person signs, and the agent later signs with that key.7 Hardt, who says he is working on OAuth 2.1, says it does not make proof of possession a required piece either.8

04

Where does it fit?

Compared withHow Proof of possession differs
DPoPDPoP is one OAuth mechanism built on proof of possession: the client signs a proof when it gets a token and a new proof with the same key when it uses it.2 RFC 7800 defines only how a token names the key.§3
HTTP Message SignaturesRFC 7800 does not sign requests itself: it puts a key in a token and leaves the way the proof travels to each protocol.§3.6 In AAuth, Hardt says, every HTTP call is signed, so the server knows the caller.8
OAuthHardt calls OAuth a framework, not a protocol, and says OAuth 2.1 does not make proof of possession a required piece.8 RFC 7800 gives a JWT a place to name its holder's key.§3
05

What goes wrong?

  • A leaked key defeats the binding. A key is only valuable if it is kept secret, and the spec says appropriate means must be used to keep private and symmetric keys from unintended parties.§4 Authors of an IETF attestation draft note that an exported key can still produce a valid proof of possession.9
  • It does not replace audience checks. The spec says applications using proof of possession should also use audience restriction, because each blocks something different: the key lets a recipient reject unauthorized senders, the audience lets it reject messages meant for another recipient.§4
  • A recipient may not understand cnf. Without an application rule, confirmation members a recipient does not understand must be ignored.§3.1 An application that needs the key checked must make sure the recipient implements those parts.§4
  • A symmetric key in a signed token would be exposed. The spec allows one as a plain "jwk" only if the JWT is encrypted; otherwise the key must be encrypted and carried as "jwe".§3.2
  • Encrypted symmetric secrets can be replayed. A signed nonce that differs for each interaction can avoid this, as can a sub-key derived for that one proof.§4
  • One key used everywhere links the holder. A key shared with many parties can serve as a correlation handle, so the spec recommends a different key for each party.§5
  • Optional binding leaves a bearer token. In one credential draft, MTAC, holder proof of possession is optional, so a credential without it can be presented by anyone holding the bytes.1
06

What should a security leader ask?

  1. 01

    Which of our access tokens and agent credentials are bearer tokens today, and which are bound to a key?

  2. 02

    Where are the private keys behind our bound tokens generated and stored, and can they be exported?

  3. 03

    Do the services that receive our tokens actually check the cnf key, or do they ignore it?

  4. 04

    Do our bound tokens also carry an audience, and does every recipient check it?

  5. 05

    Does each client use a different key for each service it talks to?

07

Sources

›Evidence: 9 claims from 9 briefings
  1. 1BI-FIND-2026-1139 · A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviourHolder proof of possession is optional in MTAC, so a credential without a holder commitment is a bearer credential that anyone holding the bytes can present. The witness co-signature guards against a compromise limited to the CA key and against dropped or truncated batches, but Burls writes that it does "NOT provide non-equivocation", and a CA and witness under one administrative authority give no independence. There is no revocation: the draft specifies none and implements none. His published verifier, aegis-verifier 0.1.3, does not implement the witness or holder checks, and the public record of batch roots is a contract on the Ethereum Sepolia testnet under a dev-grade key.
  2. 2BI-FIND-2024-0069 · In May 2023, Parecki described OAuth 2.1 as consolidation and DPoP as token protectionParecki’s May 2023 explanation of DPoP, a proof-of-possession mechanism for OAuth tokens, links token acquisition and use to the same client key. The client includes a signed proof when obtaining an access token, then makes a new proof with the same private key when using it. He described DPoP as OAuth-specific and suitable for browser JavaScript, rather than a general-purpose signing mechanism. His statement that it was near publication belongs to the talk date.
  3. 3BI-FIND-2026-0522 · Authors from Anthropic, OpenAI, Okta and Ping want the agent platform registered once, not every agentThe grant identifies the agent and does little to hold it. The assertion is a bearer credential and client authentication is optional, so whoever holds it can redeem it until it expires; proof of possession, an explicit token type and mandatory replay rejection by jti are all listed as open issues. The text never mentions revoking an agent, a platform registration or an issued token. Short lifetimes and no refresh tokens are the only brakes. What the agent may do is the server's local policy, optionally mapped from claims the platform asserts, and is declared out of scope.
  4. 4BI-FIND-2026-1027 · An agent-credential draft carries a hash of the user's instruction that its own verification steps never checkThe draft says capping a child's expiry at its parent's means revoking a parent "effectively expires all descendants, even without an active revocation lookup." Revocation never changes an expiry time, though, so the cap matters only when the parent runs out naturally. Descendants are revoked by a cascade instead: at revocation time every stored credential whose chain contains the revoked ID goes into a revocation store, and verifiers MUST check that store on every call. The delegation step checks the parent's signature, expiry and algorithm but not the store, while the human-approval path does re-check revocation. We see nothing in the text that stops a revoked but unexpired parent from minting a new child that the cascade never saw and that would pass verification. The draft defines no proof-of-possession binding, so the credentials work as bearer tokens, valid for up to 24 hours.
  5. 5BI-FIND-2026-1240 · OpenPort holds agent writes as drafts for an operator by default; its own tests check less than its spec asksThe paper specifies more than its own release tests check. At tag v0.1.0 the gate ran 20 unit tests, 80 malformed fuzz requests with no server error, and a core conformance check of 6 required endpoints. The authors' coverage table marks the state-witness check, idempotency replay and audit completeness as future work and end-to-end rate limiting as partial. The state-witness profile addresses the gap a late approval leaves, when the resource has changed since the operator looked, by re-checking state at execution and failing closed, but implementations only MAY support it. The runtime keeps state in memory and uses bearer tokens with no proof of possession, and the authors say the results "do not claim production readiness."
  6. 6BI-FIND-2026-0508 · IETF workload practices draft describes single-audience JWTs in place of stored secrets and discourages token exchangeThe draft's history says version 07 answered the Area Director by hardening its SHOULD and MUST text; version 08, posted on 30 October after more Area Director feedback, writes the exits into the rules. Section 5.1.1 now says credentials MUST be scoped as narrowly as possible unless the platform issues one credential per workload or does not let the workload influence the audience; version 07 stated the MUST flatly and conceded the exception a paragraph later. When either exception applies, deployments need to treat every party that receives the credential as able to impersonate the workload at every other party that accepts it. Where the platform or the relying party does not support proof of possession, deployments MUST apply compensating controls "such as" short-lived tokens, narrow audiences or network controls, where version 07 required all three. Tokens must not outlive their workload unless the platform cannot tell its lifetime.
  7. 7BI-FIND-2026-0415 · Google's AP2 v0.2 lets a person sign spending limits once and leave; the agent then signs each purchase itselfThe new human-not-present mode is delegation by key binding. The person signs an open mandate on a Trusted Surface that carries spending constraints and the agent's public key as a cnf claim, with a short expiry recommended; the agent later signs the closed Checkout and Payment Mandates with that key, and it may not present a second open mandate until it holds a rejection receipt for the first. That is the same move as proof-of-possession tokens in workload identity, applied to a shopping agent, and it stops exactly where our corpus's other delegation specs stop: agent-to-agent delegation is out of scope, and any attenuation beyond the constraints the person wrote down is future work.
  8. 8BI-FIND-2026-1523 · AAuth carries the signals an agent's approval needs, Hardt says; the brain that decides does not exist yetHardt, who says he is working on OAuth 2.1, says it adds no new features and does not make proof of possession a required piece either. On the web, he says, TLS lets a client know which server it reached, but "the server has no idea who's calling it". He traces the wider problem to OAuth being "a framework as opposed to a protocol", with no conformance on what a client must do to work with a given server. In AAuth every HTTP call is signed, so the server knows the caller, and he says the same signing could give webhooks a standard alternative to the shared secrets they rely on today.
  9. 9BI-FIND-2026-1733 · Attestation shows which image is running, not which copy; a new IETF draft weighs two ways to tell copies apartBoth options rest on a property this draft does not supply: assurance that the workload's key was generated inside the attested environment and cannot leave it. Neither carrying the public key in the Evidence nor certifying it shows that, and the authors note that "an exported key can still produce a valid proof of possession." They hand the property to draft-reddy-rats-key-binding, a draft that Reddy and Mihalcea co-wrote, which conveys the key in the Evidence with attributes asserting that it was locally generated and is non-exportable. This draft then gives two ways those properties can be established: as assertions by an Attesting Environment that can also act as a key store, or as an inference the Verifier may draw from endorsements for the workload's measurements. Their security section adds that a compromised attestation key can sign Evidence for any key and any state, and that no binding at the TLS layer would detect it. Binding the Evidence to the TLS connection, which stops relay, is a separate property they leave to the SEAT solution documents; in their survey, Confidential Space already derives its attestation nonce from the TLS exporter.

›In our briefings

What our briefings say about Proof of possession itself, with the most direct claims first.

  1. McGuinness's July draft requires an attester to mint an agent_instance_id that stays stable for the instance's lifetime and is independent of its proof-of-possession key. Access tokens put that identity in act.sub for delegated tokens or top-level sub for self-acting tokens, with both ai_agent and client_instance classifications. Identity survives key rotation, but refresh tokens remain bound to their issuing key: a rotated-key instance needs a fresh grant or token exchange.

    A July OAuth draft keeps agent identity stable as keys rotateBI-FIND-2026-1633

  2. Every token in Asor's chain names its holder's key in a cnf (confirmation) claim, and step 6 of the verification algorithm asks for a DPoP proof of possession only for the key in the last token. The draft says a captured token is therefore unusable without its private key; the earlier holders' keys are not proven when the chain is presented. -03 leaves mTLS-bound tokens to a later revision, so DPoP is the only holder binding it specifies. The first token must verify under a trusted root key, and the draft gives no rule for which key signs each token after it. How a verifier comes to trust the root key is still an open issue, which -03 splits into a binding gap and a recording gap. A verifier that is online and has an authenticated source of binding evidence may relieve the first, though "Network access alone establishes nothing"; the second "is not relieved by being online". For the recording gap Asor points to the retention entry in Gilda's audit-record draft, which keeps the root key material beside a record, and he mandates no trust-anchor topology. In Table 1, a verifier that never sees the request cannot conclude that the presenter holds the key.

    Asor's agent-chain draft now says a chain keeps verifying after its principal leaves, until expiry or revocationBI-FIND-2026-0502

  3. Holder binding is optional in the Recommendation. Section 4.1.3 makes it RECOMMENDED that the holder give the issuer a proof-of-possession key, binding the credential to the holder in a way the verifier can check in the presentation, and says a verifier MAY use cnf to identify that key. For SD-JWT, the spec calls the key-binding JWT optional. We read this as leaving each verifier's policy to decide whether to accept a credential presented with no holder key. Our briefing on RFC 9901 found SD-JWT itself puts the same decision on the verifier, per use case.

    W3C's JOSE and COSE rules for credentials admit there is no standard way yet to fetch a JWK from a DID URLBI-FIND-2026-2225

›Cited in 50 briefings

Every briefing that discusses this term, newest first.

  1. In SD-JWT VC, the X.509 key route does not authenticate the iss claim

    Draft · · IETF Datatracker · Oliver Terbu et al. · Aug 31, 2026

  2. Qwixl's agent-messaging draft makes data misuse attributable to a key, not preventable

    Draft · · IETF Datatracker · Luke Daniel Chapman, Qwixl · Nov 16, 2026

  3. IETF draft: AI agents you need to audit or revoke separately should not share one identity

    Draft · · IETF Datatracker · Joseph A. Salowey et al. · Jul 6, 2026

  4. AAuth gives every agent its own key, then keeps the delegation chain out of the token on purpose

    Draft · · IETF Datatracker · Dick Hardt, Hellō · Sep 25, 2026

  5. 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

  6. Google's AP2 v0.2 ties an agent's purchase to the merchant's own signed checkout by hash

    Article · · ap2-protocol.org · Google · Apr 28, 2026

  7. W3C's JOSE and COSE rules for credentials admit there is no standard way yet to fetch a JWK from a DID URL

    Article · · W3C · Michael Jones et al. · Oct 13, 2026

  8. Register a platform's identity provider once, not each agent: trust rests on its issuer, with no revocation

    Draft · · IETF Datatracker · Paul Carleton et al. · Sep 22, 2026

  9. 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

  10. JPMorgan won't let some agent sessions read and send email at once, says its security architect

    Podcast · · Insecure Agents · "Is This Tool Call Allowed? It's Never That Simple": Michael Davis (J.P. Morgan) · Sep 9, 2026

  11. IETF workload practices draft describes single-audience JWTs in place of stored secrets and discourages token exchange

    Draft · · IETF Datatracker · Arndt Schwenkschuster et al. · Sep 22, 2026

  12. 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

  13. Niyikiza's agent tokens narrow to a tool's arguments offline; revision -02 breaks -01 implementations

    Draft · · IETF Datatracker · Niki Aimable Niyikiza · Jun 15, 2026

  14. Attestation shows which image is running, not which copy; a new IETF draft weighs two ways to tell copies apart

    Draft · · IETF Datatracker · Tirumaleswar Reddy.K et al. · Oct 28, 2026

  15. The SCITT API heading to RFC leaves who may register a statement to each operator

    Draft · · IETF Datatracker · Henk Birkholz et al. · Jun 26, 2026

  16. AAuth carries the signals an agent's approval needs, Hardt says; the brain that decides does not exist yet

    Podcast · · Collateral Dialogue · Self-Driving Software: What OAuth Can’t Do for AI Agents | Dick Hardt, Hello · Sep 2, 2026

  17. An IETF draft ties attestation evidence to the certificate key, but only for the moment of the check

    Draft · · IETF Datatracker · Tirumaleswar Reddy.K et al. · Aug 15, 2026

  18. An OAuth profile for AI agents makes every hand-off cut privileges, but a shorter expiry alone counts as a cut

    Draft · · IETF Datatracker · Angel Cruz · Feb 7, 2026

  19. An end-to-end encrypted draft would hide an AI agent's task from the broker, but not who sent it or to whom

    Draft · · IETF Datatracker · Hamdy Atakora, Imara Labs · Sep 21, 2026

  20. An attested OAuth client is one somebody vouched for; revision 12 says to trust its claims only as far as the voucher

    Draft · · IETF Datatracker · Tobias Looker et al. · Oct 13, 2026

  21. A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviour

    Draft · · IETF Datatracker · Reuben Burls · Jul 21, 2026

  22. The signing-key draft under AAuth offers eight ways to trust a key, and only one says who the agent works for

    Draft · · IETF Datatracker · Dick Hardt et al. · Sep 13, 2026

  23. NIST and CISA advise one-hour tokens, 90-day keys and hardware key storage; AI agents get one paragraph

    Article · · NIST (CSRC) · Ryan Galluzzo et al. · Sep 15, 2026

  24. Coetzee's agent token is good for one declared call, and the hard part is hashing that call the same way twice

    Draft · · IETF Datatracker · Rudolf Jacobus Coetzee · Jul 19, 2026

  25. The IETF's Workload Proof Token proves the key but leaves major portions of the request unprotected, and the draft says so itself

    Draft · · IETF Datatracker · Brian Campbell et al. · Aug 27, 2026

  26. The 'intent' in the Agentic JWT draft is a workflow step, and the agent reports its own fingerprint

    Draft · · IETF Datatracker · Abhishek Goswami · Jul 1, 2026

  27. IPC passes a delegation chain between services; its signature proves only the last signer

    Draft · · IETF Datatracker · Ravi Kant Sharma, Ericsson · Jul 23, 2026

  28. A July OAuth draft keeps agent identity stable as keys rotate

    Draft · · IETF Datatracker · Karl McGuinness · Jul 5, 2026

  29. 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

  30. An agent-credential draft carries a hash of the user's instruction that its own verification steps never check

    Draft · · IETF Datatracker · Chudah Yakung, Attest · Mar 26, 2026

  31. A cross-domain agent trust draft requires the call chain, then leaves its format to a version that never came

    Draft · · IETF Datatracker · Daniel King et al. · Mar 15, 2026

  32. A Huawei agent draft keeps one 'master agent' as gatekeeper and drops the line that had servers re-check it

    Draft · · IETF Datatracker · Yuan Ni et al. · Feb 28, 2026

  33. OpenPort holds agent writes as drafts for an operator by default; its own tests check less than its spec asks

    Paper · · arXiv · Genliang Zhu et al. · Feb 22, 2026

  34. A Biscuit holder can narrow a token offline, and widening it takes an explicit trust rule

    Article · · Eclipse Biscuit · Eclipse Biscuit project · Oct 21, 2025

  35. Attested workload identities need a mapping that survives authorized updates

    Draft · · IETF Datatracker · Mark Novak et al. · Oct 20, 2025

  36. WIMSE leaves its application proof choice open, with different integrity coverage

    Draft · · IETF Datatracker · Brian Campbell et al. · Oct 16, 2025

  37. IETF HSM evidence draft: a device may call a key device-bound only if no route, now or later, can export it

    Draft · · IETF Datatracker · Jean-Pierre Fiset et al. · Oct 10, 2025

  38. Agentic JWT binds an agent's token to its task and its own code, but it is one author's draft, not a standard

    Paper · · arXiv (cs.CR preprint 2509.13597) · Abhishek Goswami · Sep 16, 2025

  39. OpenID's credential issuance standard is OAuth for wallets, and its recommended wallet check is built for native apps

    Article · · OpenID Foundation · Torsten Lodderstedt et al. · Sep 16, 2025

  40. Attestation-key EKU draft makes the CA ensure only the Attester controls the key

    Draft · · IETF Datatracker · Jean-Pierre Fiset et al. · Aug 5, 2025

  41. CMP adds KEM authentication with fresh secrets for each certificate-management operation

    Article · · RFC Editor · Hendrik Brockhaus et al. · Jul 29, 2025

  42. China Mobile's agent-identity draft had one MUST, binding each agent to its user, and no mechanism to do it

    Draft · · IETF Datatracker · Kehan Yao et al. · Jul 1, 2025

  43. RFC 9711 standardizes the attestation token and sets no security floor for the device that signs it

    Article · · RFC Editor · Laurence Lundblade et al. · Apr 2025

  44. RFC 9679 lets a CBOR token name its holder's key by thumbprint alone, and leaves finding the key to the verifier

    Article · · RFC Editor · Kohei Isobe et al. · Dec 20, 2024

  45. In GNAP (RFC 9635), a client is known by its key, and tokens stay key-bound unless the server flags them bearer

    Article · · RFC Editor · Justin Richer et al. · Oct 1, 2024

  46. In May 2023, Parecki described OAuth 2.1 as consolidation and DPoP as token protection

    Video · · Identiverse - A CRA Resource · Jun 17, 2024

  47. COSE certificate headers leave trust validation with the relying application

    Article · · RFC Editor · Jim Schaad · Feb 16, 2023

  48. RFC 7800 gives a JWT one place to name its holder's key, and leaves the proof to each protocol

    Article · · RFC Editor · Michael B. Jones et al. · Apr 1, 2016

  49. RFC 7644 leaves SCIM's access rules to each deployment, and a bulk request can succeed in part

    Article · · RFC Editor · Phil Hunt et al. · Sep 1, 2015

  50. A 2025 workload identity draft kept credential exchange for crossing trust domains, and never to gain more trust

    Draft · · IETF Datatracker · Arndt Schwenkschuster, Defakto · Mar 3, 2025

Related terms

What appears alongside Proof of possession in the same briefings.

HTTP Message SignaturesOAuth

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.

Follow us on LinkedInRSS feed

Working on an identity problem that involves Proof of possession? Talk to Boston Identity