]>

kieran@kierans.net

sec WIMSE AI agents delegation OAuth WIMSE capability Autonomous AI agents increasingly require access to protected resources across multiple service providers on behalf of human users. Existing OAuth 2.0 extensions address individual aspects of this problem (token exchange, proof-of-possession, and structured authorization) but no current specification defines how these mechanisms compose into a coherent credential delegation framework for AI agents. This document specifies the Credential Delegation Protocol: a profile of OAuth 2.0 Token Exchange (RFC 8693), Demonstrating Proof-of-Possession (RFC 9449), Rich Authorization Requests (RFC 9396), and Client-Initiated Backchannel Authentication (OpenID Connect CIBA) that enables human users to delegate scoped, attenuated credentials to AI agents operating across heterogeneous service providers. The protocol defines: agent identity lifecycle management using ephemeral key pairs; capability-shaped delegation tokens bound to specific operations and resources; credential wrapping semantics that prevent exposure of underlying OAuth tokens to agents; consent-gated delegation flows for asynchronous agents; real-time cascading revocation; and tamper-evident audit chains. This document does not define new token formats, new OAuth grant types, or modifications to existing authorization server behavior. It specifies how existing mechanisms are combined to achieve secure, auditable credential delegation for AI agents. Status information for this document may be found at . Discussion of this document takes place on the Workload Identity in Multi System Environments (WIMSE) Working Group mailing list (), which is archived at . Subscribe at . Source for this draft and an issue tracker can be found at .

Introduction

Problem Statement OAuth 2.0 solved human-to-service authorization. When a user authorizes an application to act on their behalf, the application receives an access token representing that delegation. This model assumes a human present at a browser for the consent ceremony. AI agents operate autonomously. They are ephemeral (spawned on demand), numerous (many agents per user per service), and adversarially promptable: a compromised prompt can direct an agent to misuse any credential it holds. Existing standards fail the agent delegation use case in seven specific ways: No agent identity primitive. OAuth clients require pre-registration. Agents are ephemeral and cannot register at instantiation time. SPIFFE requires admin provisioning. No standard defines bootstrapping an agent identity from nothing. No delegation chain attenuation. When Agent A sub-delegates to Agent B, existing standards do not enforce that authority can only narrow. RFC 8693 records delegation chains via nested act claims but treats them as "informational only": no enforcement, no structural guarantee. The delegation chain splicing vulnerability (disclosed to the OAuth WG mailing list, February 26, 2026 ) demonstrates that Sections 2.1-2.2 of permit a compromised intermediary to present mismatched subject_token and actor_token from different delegation contexts, producing a properly-signed token asserting a delegation chain that never occurred. No credential wrapping. Agents need to use credentials at resource servers that do not understand delegation. No standard defines how a delegation token authorizes credential exercise without exposing the raw credential to the agent. No granular authorization. OAuth scopes are coarse string identifiers. Agents need resource- and operation-level capability binding: not "can access Google Drive" but "can read file X in folder Y until time T." RFC 9396 provides the structural mechanism but no agent-specific vocabulary or attenuation rules. No synchronous revocation cascade. Revoking a root delegation must immediately invalidate all derived delegations. UCAN revocation is gossip-based. RFC 8693 explicitly defers revocation to implementations. No existing standard provides sub-second cascading revocation for a delegation tree. No asynchronous consent. Agent-initiated flows require user approval without a redirect URI. CIBA provides the mechanism but is not profiled for agent delegation scenarios, and CIBA + DPoP interaction is underspecified. No delegation audit chain. No standard defines an immutable, portable audit trail for "Agent A used User B's credential C to perform operation D at time T via delegation chain E." This document profiles existing standards to address all seven gaps.

Relationship to Existing Work The IETF landscape for agent authorization moved quickly during 2026. The OAuth Working Group was rechartered in June 2026 with agents acting on behalf of users explicitly in scope. The WIMSE Working Group held an interim meeting on June 3, 2026 dedicated to AI agent authentication and authorization and is discussing adoption of that work. Several dozen individual drafts now address some slice of agent identity or delegation. This section positions this protocol against the working group documents and the individual drafts closest to its mechanics. It is not a survey.

Frameworks and Working Group Documents draft-klrc-aiagent-auth (): A framework for AI agent authentication and authorization, at -02 with -03 in preparation, and the subject of the WIMSE adoption discussion noted above. The framework identifies the need for concrete delegation mechanics and defers them to OAuth flows. This document supplies those mechanics and is designed as a companion, not a replacement. draft-ietf-wimse-arch (): Defines the trust domain model and terminology assumed here for WIMSE deployments. The Delegation Server is a service within a trust domain, and Agent identities established per are compatible with WIMSE workload identifiers. draft-ietf-wimse-wpt (): Defines the Workload Proof Token (WPT), a proof-of-possession mechanism for workloads. This protocol binds Delegation Tokens with DPoP instead: DTs are exercised over plain HTTP by ephemeral agents that may sit outside any WIMSE trust domain, the deployment shape DPoP already serves. A future revision may profile WPT for DS-to-resource-server hops inside WIMSE trust domains. draft-ni-wimse-ai-agent-identity (): Addresses agent identity within WIMSE. The did:key model used here is compatible. This document adds credential wrapping, revocation, consent gating, and chain verification. draft-reece-wimse-cross-org-delegation (): A problem statement and requirements catalogue for cross-organizational delegation, used on the WIMSE list as a common frame for evaluating delegation proposals. A future revision of this document will map the protocol against those requirements.

Delegation Chain Mechanics draft-ietf-oauth-identity-chaining (): Approved for publication in 2026. Chains identity and authorization across trust domains by exchanging a token in one domain for a JWT assertion accepted in another. It preserves identity across hops but does not constrain what each hop may do. This protocol adds structural attenuation and receipt-based chain verification on the same RFC 8693 substrate. The two compose. draft-niyikiza-oauth-attenuating-agent-tokens (): Defines tokens a holder can attenuate offline, in the macaroon tradition. Same goal as , different trust model: AAT verification relies on the root issuer key and offline derivation, while this protocol keeps the Delegation Server in the loop at every hop in exchange for synchronous revocation () and server-checked strict-subset validation. draft-gco-oauth-delegate-sd-jwt (): Delegates SD-JWT credentials holder-to-holder by allowing a Key Binding JWT to act as an SD-JWT. It operates at the credential format layer; this protocol operates at the delegation service layer. A DS could issue capability attestations as Delegate SD-JWTs without changing the mechanics defined here. draft-zhu-oauth-async-delegation (): Defines delegated refresh tokens so agents can act while the user is offline. This protocol answers the same need by keeping refresh tokens in the Credential Vault and giving agents only short-lived DTs. The two are alternative positions on whether agents may hold long-lived credentials at all. draft-oauth-ai-agents-on-behalf-of-user (; expired February 2026): Introduced requested_actor and actor_token parameters binding the acting agent into the authorization code exchange. That closes the same actor-binding weakness this protocol closes with prf receipts, but only for the first hop and only at the authorization server. See .

Credential Intermediation draft-hartman-credential-broker-4-agents (): Specifies a credential vaulting broker that mediates agent API access through short-lived, narrowly scoped proxy credentials. Architecturally the closest work to the Delegation Server and Credential Vault defined here. CB4A defines the broker; this document additionally defines the delegation chain (attenuation, receipts, sub-delegation), the consent flow, and the revocation cascade such a broker must enforce. draft-araut-oauth-transaction-tokens-for-agents (): Extends transaction tokens to carry agent context in the act claim for intra-domain call chains. Complementary: transaction tokens propagate context within a domain after authorization exists; this protocol governs how the authorization comes to exist. draft-schwenkschuster-wimse-credential-exchange (; expired): Despite the similar name, addresses a workload exchanging its own credential for a different format or trust domain. This document addresses human-to-agent delegation over stored credentials that are never issued to the requesting party.

Consent, Mandates, and Audit Evidence draft-yossif-agent-mandate-problem (): A problem statement on verifiable human mandates: intent authorized at one time, executed autonomously later. The consent records () and capability constraints defined here are a concrete answer to part of that problem space. draft-nelson-agent-delegation-receipts (): Defines user-signed delegation receipts anchored to an append-only log before an agent acts, removing the operator as a trusted intermediary. The prf receipts in are DS-signed and bind hops within a chain; Nelson's receipts are user-signed and bind the user's instruction to the run. They answer different trust questions and can coexist. draft-nennemann-wimse-ect (): Execution Context Tokens define an audit record format. This protocol's audit chain is designed to be ECT-compatible. draft-goswami-agentic-jwt (; expired July 2026): Proposed agent checksums and intent binding. The agent_checksum claim () remains defined for compatibility, with no dependency taken. A US patent has been filed on the underlying mechanism; this specification avoids the patented specifics.

Design Principles Compose, don't invent. Every mechanism reuses an existing standard. New concepts appear only where the gaps identified in are confirmed. Attenuation is structural. Agents cannot widen authority at any delegation hop. This is enforced by the Delegation Server, not by trusting agents. Inspired by UCAN attenuation semantics and the object-capability model . Credentials are never possessed. Agents receive and exercise delegated authority through the Delegation Server. Raw credentials never cross the boundary to the agent host. This architecturally eliminates the confused deputy attack surface for credential theft . Revocation is synchronous. A user revoking delegation takes effect within the SLA defined in . Eventual consistency is not acceptable for credential revocation in adversarially-promptable systems. Capability-shaped, not identity-scoped. Delegation tokens authorize specific operations on specific resources with specific constraints, not "Agent X can access Service Y." Follows the NORA (designation = authority) principle from capability security. Chain integrity is cryptographic. Each delegation hop produces a signed receipt. The chain is verifiable end-to-end without trusting intermediate agents, addressing the RFC 8693 delegation chain splicing vulnerability.

Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Delegation Server (DS):

A service that issues, manages, and revokes Delegation Tokens on behalf of Subjects. The DS maintains the Credential Vault and enforces delegation policies. It acts as an intermediary between Agents and upstream OAuth authorization servers.

Agent:

An autonomous software entity that performs actions on behalf of a Subject across one or more service providers. An Agent authenticates to the DS using an ephemeral key pair and receives Delegation Tokens authorizing specific operations. An Agent MUST NOT possess or have access to the underlying credentials stored in the Credential Vault.

Subject:

The human user who authorizes credential delegation. The Subject authenticates to the DS, deposits credentials into the Credential Vault, defines delegation policies, and approves or denies consent-gated delegation requests.

Delegation Token (DT):

A signed JWT issued by the DS that authorizes an Agent to perform specified operations on specified resources via the DS. A DT contains capability claims structured per , a DPoP key binding for proof-of-possession verification, and an opaque credential handle. Delegation Tokens MUST NOT contain raw OAuth tokens or credentials.

Credential Vault:

Server-side secure storage maintained by the DS that holds OAuth tokens and credentials deposited by Subjects. Credentials are referenced by opaque handles and exercised exclusively by the DS on behalf of Agents presenting valid Delegation Tokens.

Capability:

A structured authorization grant specifying a permitted operation, a target resource, and optional constraints (time bounds, rate limits, argument restrictions). Expressed using the authorization_details object defined in and bound to a specific Delegation Token.

Attenuation:

The process by which a Capability is further constrained when delegated from one entity to another. An attenuated Capability MUST be a strict subset of its parent. Authority can only narrow at each delegation hop; it can never widen.

Delegation Chain:

An ordered sequence of Delegation Tokens from Subject to Agent_1 to Agent_2 ... to Agent_N, where each link attenuates the authority of the previous. The chain is verifiable from any link back to the root consent event via signed delegation receipts.

Credential Exercise:

The act of the DS using a stored credential on behalf of an Agent. The Agent presents a valid DT and DPoP proof to the DS, which validates the DT, retrieves the credential from the Vault, calls the resource server, and returns only the API response.

Architecture Overview

Retrieves cred -> Calls RS | | | | Returns API response (never raw credential) | | | +----------------------------------------------+ | | | | +----------------------------------------------+ | | | Revocation & Audit | | | | Synchronous cascade | Tamper-evident log | | | +----------------------------------------------+ | +--------------------------+---------------------------+ | v +------------------------------------------------------+ | AGENT | | | | 1. Generates ephemeral did:key | | 2. Authenticates via JWT Bearer (RFC 7523) | | 3. Requests delegation (capabilities + constraints) | | 4. Receives DT (DPoP-bound, capability-shaped) | | 5. Exercises credentials via DS /exercise endpoint | | 6. Sub-delegates via attenuation (optional) | | | | TRUST BOUNDARY: Agent sees DT + API responses only | | Agent NEVER sees: raw tokens, refresh tokens, keys | +------------------------------------------------------+ ]]>

Critical architectural property: The Agent's trust boundary extends only to the network interface of the Delegation Server. The Agent never crosses into the Vault. This eliminates credential theft as an attack surface: there is nothing for a compromised agent to exfiltrate.

Protocol Version Negotiation The canonical protocol version for this draft is 0.1.0. HTTP protocol clients MUST advertise the highest Cred Protocol version they support for a request using the Cred-Protocol-Version request header. The header value MUST be a single semantic-version string in MAJOR.MINOR.PATCH form. Delegation Servers MUST set Cred-Protocol-Version on every protocol response, including error responses, to the version selected by the server for that response. Each implementation MUST maintain: a supported-version set, ordered from newest to oldest a configured version floor, below which requests are rejected For this draft, the supported-version set is ["0.1.0"] and the default version floor is 0.1.0. A Delegation Server MUST reject a request that advertises a version lower than the configured floor or a version outside the supported-version set. The server MUST return HTTP 426 with a JSON body containing:

The response MUST also include Cred-Protocol-Version set to the server's current version so clients can distinguish protocol drift from authentication, authorization, or policy failures. During the 0.1.0 compatibility window, a missing Cred-Protocol-Version request header MAY be treated as 0.1.0. This allowance exists only because there is no earlier wire version to downgrade to. Implementations that raise the version floor above 0.1.0 MUST reject missing protocol-version headers with protocol_version_unsupported.

Agent Identity

Ephemeral Key Pairs An Agent MUST generate an Ed25519 or P-256 key pair at instantiation. The DID is derived deterministically from the public key using the did:key method , a DID method conforming to . No pre-registration is required. The DS MUST NOT reject a DID it has not seen before.

Agent Authentication The Agent authenticates to the DS using a JWT Bearer assertion signed with the private key corresponding to its did:key. Required claims: Claim Value Notes iss Agent DID did:key:z... sub Agent DID Same as iss aud DS endpoint URL   iat Current time   exp iat + max 300s 5-minute maximum jti Unique nonce Prevents replay Optional claims: Claim Value Notes agent_model Model identifier Vendor-prefixed, e.g., "vendor:model-id" agent_operator Operator DID or URL Organization running the agent agent_checksum SHA-256 hash Intent binding per ; advisory only due to IPR

Agent Lifecycle Agents SHOULD generate a new key pair per session. The DS MAY require Subject pre-authorization of specific agent DIDs or agent operators before issuing Delegation Tokens. When an agent terminates, the DS SHOULD invalidate any active Delegation Tokens bound to that agent's DPoP key within the revocation SLA.

Delegation Token Format A Delegation Token is a DPoP-bound JWT issued by the Delegation Server.

Required Claims Claim Value Specification iss DS identifier   sub Subject identifier User on whose behalf delegation occurs act {"sub": "<Agent DID>"} Per RFC 8693, Section 4.1 authorization_details Capability array Per RFC 9396 cnf {"jkt": "<DPoP thumbprint>"} Per RFC 9449 iat Issuance time   exp Expiry time Max 1 hour; 15 minutes RECOMMENDED jti Unique identifier   consent_id Consent record ID Traceable to root consent event credential_handle Opaque string References Vault entry; MUST NOT be the credential itself

Capability Structure Capabilities are expressed as RFC 9396 authorization_details objects:

The DS MUST validate that capabilities in a sub-delegation request are a strict subset of the parent DT's capabilities. Validation is the DS's responsibility, not the requesting agent's.

Delegation Chain Integrity To address the delegation chain splicing vulnerability in Sections 2.1-2.2 of , each delegation hop MUST produce a signed delegation receipt containing: the parent DT's jti, the child DT's jti, the delegating agent's DID, the receiving agent's DID, and the attenuated capability set. The receipt is signed by the DS and included in the child DT as the prf claim (an array of receipt content identifiers). Chain verification traces prf links back to the root consent event.

Credential Wrapping

Exercise Flow The Agent presents a valid DT and DPoP proof to the DS's /exercise endpoint. The DS: Validates the DT signature and expiry Verifies the DPoP binding (htm, htu, nonce) Checks that the requested operation is within the DT's authorization_details Retrieves the credential from the Vault using the opaque credential_handle Performs the authorized API call against the resource server using the stored credential Returns only the API response The raw credential MUST NOT appear in any agent-facing response.

Proxy Semantics For resource servers that accept standard OAuth Bearer tokens: the DS acts as a reverse proxy, performing the actual API call with the stored credential. The Agent's HTTP request to /exercise specifies the operation and parameters; the DS maps these to the upstream API call.

Native Resource Server Support (Future) For resource servers that natively support this protocol: the DS performs RFC 8693 token exchange, issuing a scoped access token containing the DT's act claim and authorization_details for direct presentation at the resource server. This eliminates the proxy step for participating services.

Revocation When a Subject revokes a delegation (root or any subtree node), the DS MUST: Invalidate the specified DT within 1 second Cascade revocation to all descendant DTs within 5 seconds Return HTTP 401 with error: delegation_revoked for any in-flight /exercise request using a revoked DT Write an immutable revocation event to the audit log with a revocation_time claim Revocation endpoint: DELETE /delegation/{jti} The response MUST include a revoked_count field indicating the number of tokens invalidated in the cascade.

Consent Flow

CIBA-Derived Agent Consent When an Agent requests a delegation that exceeds pre-authorized policies, the DS initiates a CIBA backchannel authentication request to the Subject. The Subject approves or denies on their registered device. On approval, the DS issues the Delegation Token. The Agent polls or receives a push notification when the token is available.

Consent Records The DS MUST store an immutable record of each consent event, including: Subject identifier, Agent DID, capabilities granted, grant time, expiry, and the full delegation chain context at time of consent. Records MUST be retained for at least 90 days.

Re-Consent Triggers Consent MUST be re-requested when: an Agent requests broader capabilities than previously granted the grant has expired the agent_operator claim changes a security event triggers policy re-evaluation

Security Considerations

Confused Deputy Mitigation Capability-shaped tokens eliminate ambient authority. An adversarially-prompted agent cannot widen its own capabilities since attenuation is DS-enforced. OAuth access tokens are ambient authority: any code holding the token exercises its full scope. For adversarially-promptable agents, this is a structural exploit vector. Cred delegation tokens are operation-bound, resource-specific, and constraint-bearing, following the principle that designation should equal authority .

Delegation Chain Splicing The mandatory prf chain with DS-signed receipts addresses the vulnerability in Sections 2.1-2.2 of disclosed to the OAuth WG on February 26, 2026 . Each receipt cross-references parent and child DT jti values along with both agent DIDs, preventing presentation of tokens from mismatched delegation contexts. The requested_actor/actor_token binding proposed in mitigates the same class of attack at the authorization server for the first delegation hop; the prf chain extends equivalent protection across every subsequent hop (see ). The two can compose: a DS can accept an actor-bound access token issued under as the credential it wraps, then produce prf receipts for every sub-delegation beyond that first hop.

DPoP Binding DPoP prevents Delegation Token theft and replay by requiring cryptographic proof of private key possession on every /exercise request. The proof is bound to the htm (HTTP method) and htu (URI) of the specific request, preventing re-use across operations.

Prompt Injection Containment Because credentials never reach the agent host, a successful prompt injection attack cannot exfiltrate credentials. The agent can only exercise capabilities already granted in its DT, and only via the DS /exercise endpoint. The blast radius of a compromised agent is bounded by its current DT's authorization_details.

Chain Depth Limits Implementations SHOULD enforce a maximum delegation chain depth of 5. Unbounded sub-delegation creates exponential revocation cascades and audit complexity.

Token Lifetime The default DT lifetime of 15 minutes limits the blast radius of any single token compromise. Implementations MUST NOT issue DTs with a lifetime exceeding 1 hour.

Model Identity and Substitution The agent_model and agent_checksum claims are self-asserted. A credential valid for an agent remains valid if the operator substitutes the underlying model, a threat class raised against the klrc framework in March 2026. Verifying model identity requires attestation of the agent runtime and is out of scope for this document. Deployments that need it should treat these claims as advisory policy input, not proof, and look to remote attestation work in the RATS WG.

IANA Considerations This document requests registration of: cred_delegation as an authorization_details type in the OAuth Authorization Server Metadata registry (per ) agent_model, agent_operator, consent_id, credential_handle, prf as JWT claim names in the JSON Web Token Claims registry urn:ietf:params:oauth:grant-type:cred-delegation as a URI in the OAuth Parameters registry

In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. The OAuth 2.0 authorization framework enables a third-party application to obtain limited access to an HTTP service, either on behalf of a resource owner by orchestrating an approval interaction between the resource owner and the HTTP service, or by allowing the third-party application to obtain access on its own behalf. This specification replaces and obsoletes the OAuth 1.0 protocol described in RFC 5849. [STANDARDS-TRACK] This specification defines a protocol for an HTTP- and JSON-based Security Token Service (STS) by defining how to request and obtain security tokens from OAuth 2.0 authorization servers, including security tokens employing impersonation and delegation. This document describes a mechanism for sender-constraining OAuth 2.0 tokens via a proof-of-possession mechanism on the application level. This mechanism allows for the detection of replay attacks with access and refresh tokens. This document specifies a new parameter authorization_details that is used to carry fine-grained authorization data in OAuth messages. This specification defines the use of a JSON Web Token (JWT) Bearer Token as a means for requesting an OAuth 2.0 access token as well as for client authentication. OpenID Foundation RC v1.0 ACM SIGOPS Operating Systems Review PhD thesis, Johns Hopkins University Work in Progress, draft-klrc-aiagent-auth-02 IETF OAuth Working Group draft-ietf-oauth-identity-chaining-15, approved for publication Work in Progress, draft-niyikiza-oauth-attenuating-agent-tokens-01 Work in Progress, draft-gco-oauth-delegate-sd-jwt-00 Work in Progress, draft-zhu-oauth-async-delegation-04 Work in Progress, draft-hartman-credential-broker-4-agents-00 Work in Progress, draft-araut-oauth-transaction-tokens-for-agents Work in Progress, draft-nelson-agent-delegation-receipts-05 Work in Progress, draft-reece-wimse-cross-org-delegation-00 Work in Progress, draft-yossif-agent-mandate-problem-00 OAuth Working Group mailing list Work in Progress (expired), draft-oauth-ai-agents-on-behalf-of-user-02 IETF WIMSE Working Group Work in Progress, draft-ietf-wimse-arch-08 IETF WIMSE Working Group Work in Progress, draft-ietf-wimse-wpt-01 Work in Progress, draft-ni-wimse-ai-agent-identity-02 Work in Progress (expired), draft-schwenkschuster-wimse-credential-exchange-03 Work in Progress (expired July 2026), draft-goswami-agentic-jwt-00 Work in Progress, draft-nennemann-wimse-ect-00 OWASP NIST NCCoE

Acknowledgments The design of this protocol draws on UCAN attenuation semantics , the object-capability literature, and ongoing work in the IETF WIMSE and OAuth working groups. Threat modeling was informed by and .