EMILIA Protocol, Inc.

US team@emiliaprotocol.ai

sec AI agents evidence challenge step-up When a relying party refuses a consequential agent action because authorization evidence is missing, stale, or unverifiable, the agent needs a machine-readable description of what remains necessary. This document defines a transport-neutral Authorization Evidence Challenge data model bound to the relying party's exact action. The challenge identifies outstanding evidence requirements, freshness and status constraints, acceptable presentation profiles, and retry state. It authorizes nothing, transfers no admission ownership, and provides no promise that a later request will execute. The document also defines an HTTP binding using 403 Forbidden and RFC 9457 Problem Details, and describes an informative gateway-handoff profile for DMSC-style federation. The gateway profile communicates evidence requirements; it does not solve conserved admission or double-admission across independently operated gateways. A challenge can synchronize corrected retries and amplify load. The core therefore defines optional retry timing with per-challenge jitter, and the HTTP binding maps its lower bound to Retry-After. Retry timing controls presentation attempts only; it does not authorize the action or make an uncertain action safe to repeat.

Introduction Evidence formats describe artifacts such as receipts, permits, attestations, and logs. They do not by themselves define the live interaction that follows when a relying party finds the presented evidence insufficient. Without a common challenge, each agent protocol invents its own vocabulary for missing evidence, acceptable presentations, freshness, status checking, and retry. This document defines the challenge as an application data model, independent of HTTP, a particular agent protocol, and any one evidence format. The nearest prior art is OAuth step-up authentication , which lets a resource identify stronger authentication requirements. AE-CHALLENGE applies the interaction shape to authorization evidence about an exact proposed action. A challenge is a refusal with information. It is not an authorization decision, reservation, capability, ownership transfer, or promise of execution. Satisfying it causes the relying party to evaluate a new presentation under its live local policy.

Requirements Language 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.

Terminology

Relying party:

The component that decides whether evidence satisfies its requirements for a proposed action.

Presenter:

The agent, gateway, or other component that receives a challenge and may later present evidence.

Exact action:

The relying-party-derived material action that would be admitted, including every field its selected action profile treats as consequential.

Challenge:

A single-use, expiring description of evidence still required for one exact action.

Transport-Neutral Core Data Model The AE-CHALLENGE-v1 data model is independent of its carrier. A transport binding MUST preserve every member that the relying party uses when registering and later consuming the challenge. A binding MUST also provide authenticated integrity and peer identification, or carry the challenge in an authenticated envelope that provides those properties. The JSON format defined by is the reference serialization of the core model. A non-JSON binding MUST define an unambiguous mapping to the same members and value types. ", "action_digest": "sha256:", "action_profile": "https://example.net/action/payment-v3", "audience": "https://gateway-b.example", "policy_id": "https://gateway-b.example/policies/high-risk-v4", "policy_digest": "sha256:", "required_evidence": [ { "requirement_id": "human-approval", "type": "https://example.net/e/authorization-receipt", "profiles": ["https://example.net/profiles/receipt-v1"], "max_age_sec": 300, "status": "current" } ], "present_as": ["https://example.net/p/ep-aec-v1"], "obtain_hints": [ { "requirement_id": "human-approval", "mechanism": "https://example.net/flows/approval-v1", "uri": "https://approver.example/tasks/123" } ], "retry_timing": { "not_before": "2026-08-09T23:50:30Z", "jitter_sec": 20 }, "expires_at": "2026-08-10T00:05:00Z" } ]]>

Action and Policy Binding @version MUST equal AE-CHALLENGE-v1. A recipient that does not implement this version MUST refuse automated processing. challenge_id is an opaque correlation identifier. Authorization decisions MUST NOT depend on its global uniqueness. nonce MUST be unpredictable, non-empty, and single-use. The relying party MUST bind the complete challenge body to the durable nonce record before exposing the challenge. action_digest MUST be computed by the relying party from the exact action it would admit. It MUST NOT be copied from the presenter or from presented evidence. action_profile MUST be present, MUST be an absolute URI as defined by , and identifies the canonicalization or mapping profile used to derive the digest. It is descriptive to the presenter; it does not let the presenter select or weaken that profile. If two parties cannot establish the same material action under an understood profile, the challenge cannot repair the disagreement and the relying party MUST refuse automated admission. audience identifies the component that is expected to answer the challenge. A binding or authenticated envelope MUST let the presenter establish the challenge issuer. A relying party processing a returned challenge MUST verify the issuer and audience before using any obtain hint or disclosing evidence. policy_id and policy_digest identify the local policy state from which the requirements were derived. They do not transfer policy authority to the presenter. At evaluation time the relying party MUST use its authenticated live policy and MUST NOT use a policy supplied in the response. Delegation, translation, or task rewriting does not preserve one action binding across hops. Each receiving relying party MUST re-evaluate the action it is asked to admit under its own policy and action profile. If that relying party issues a challenge, it MUST derive a new action_digest and register its own challenge binding. A prior hop's challenge or decision record MAY be presented as evidence or context, but it MUST NOT substitute for the receiving relying party's binding to a rewritten action. A chain of per-hop records does not by itself prove that the actions at different hops are semantically equivalent.

Evidence Requirements and Presentation Each required_evidence entry MUST contain a requirement_id and type. Requirement identifiers are unique within one challenge and correlate requirements with acquisition hints and diagnostics. Every evidence type, evidence profiles entry, and present_as entry MUST be an absolute URI as defined by so that the same identifier has portable meaning across transport protocols and conformance corpora. Unless the defining specification says otherwise, these identifiers are compared as exact strings. A recipient MUST NOT infer security semantics from URI prefixes or substring matches; it uses only locally configured semantics for the complete identifier. An entry MAY contain profiles, proof_predicates, max_age_sec, and status. Profiles are alternatives acceptable for that requirement. max_age_sec is measured at the relying party's evaluation time using the issuance or observation time defined by the selected evidence profile. status equal to current requires a separately authenticated status result satisfying the relying party's freshness policy; it does not mean that absence from an unauthenticated revocation list is sufficient. present_as lists alternative presentation profiles. The order expresses the relying party's preference, not a security ranking. Supporting a presentation profile does not imply support for every evidence type named by the challenge. The presenter selects one mutually supported profile and includes enough information for the relying party to match every component to a requirement. obtain_hints are untrusted routing hints. Following a hint, authenticating to its URI, or receiving an "approved" response confers no authority at the relying party. Evidence obtained through a hint remains subject to native verification, exact-action matching, freshness and status checks, policy evaluation, and one-time consumption. Unknown transports MAY omit obtain hints entirely. Capability discovery MAY advertise evidence types, presentation profiles, or acquisition mechanisms that an implementation supports. Such discovery is not an authorization decision and does not establish that any advertised choice is sufficient for a particular action. The relying party MUST determine sufficiency using its authenticated local policy at evaluation time. A cached discovery response, a mutually supported profile, or a successful acquisition flow MUST NOT waive an unsatisfied local requirement.

Lifecycle, Replay, and Retry expires_at MUST be an absolute timestamp. The relying party MUST refuse an expired challenge before evidence or policy evaluation. Challenge expiry does not extend the validity of any evidence and evidence expiry does not extend the challenge. Evaluation is ordered: version and structure; authenticated issuer and audience; durable body binding; action agreement; expiry; atomic nonce consumption; native evidence verification; and local policy evaluation. Malformed input that cannot be associated with a registered challenge, or a presentation that does not agree with the registered exact action, is refused without consuming a nonce. Once structural and action matching succeed, the nonce MUST be consumed atomically before evidence evaluation. A concurrent duplicate that arrives after another presentation has consumed the nonce, whether that first evaluation is still in flight or has completed, MUST be refused as a replay. It MUST NOT receive a second evidence evaluation, an indeterminate result merely because the first evaluation is in flight, or a second admission. An implementation MAY distinguish in-flight from completed consumption internally for operations and audit, but that distinction does not reopen the challenge. If evidence remains missing, stale, or unverifiable, the relying party MAY issue a follow-up challenge. A follow-up MUST use a fresh challenge identifier and nonce, MUST remain bound to the same locally derived action digest and action profile, and SHOULD list only the requirements that remain unsatisfied. It is a new refusal, not a continuation of the consumed nonce. retry_timing is OPTIONAL. When present, it MUST contain a not_before Internet timestamp as defined by and MAY contain a non-negative integer jitter_sec. A presenter that processes this member MUST NOT send the challenge response before not_before. When jitter_sec is greater than zero, the presenter SHOULD add an independently selected delay between zero and that many seconds. The selected delay SHOULD vary for each challenge and MUST NOT be treated as extending expires_at. A relying party that expects many correlated refusals SHOULD issue recipient-specific or challenge-specific retry schedules rather than one common retry instant. It SHOULD provide a jitter interval wide enough for its expected population and recovery capacity. A presenter remains responsible for its own backoff and rate limits when timing is absent. A challenge MUST NOT be issued solely to signal overload when the carrying protocol has a distinct overload response. If evidence insufficiency and load pressure coexist, retry timing can pace the new evidence presentation without changing the refusal semantics. Ignoring retry_timing does not make evidence sufficient or authorize an action, but it can defeat the issuer's load-control goal. An issuer that requires the recipient to understand the timing member MUST include retry_timing in critical. A presenter that cannot establish whether not_before has passed MUST NOT infer permission to send early; it uses its local backoff policy or declines automated retry. Unknown acquisition status, timeout, transport failure, or an unverifiable response is indeterminate. It MUST NOT be interpreted as evidence satisfaction or permission to retry an action whose effect may already have occurred.

Extensibility Recipients MUST ignore an unknown member unless its name appears in the optional critical array. A recipient that does not understand every member named by critical MUST refuse automated processing. A specification defining an extension MUST state whether that extension is safe to ignore and how it is bound to the registered challenge body.

Carrier Semantics A transport binding SHOULD carry a challenge as a structured refusal or error payload when the carrying protocol provides such a facility. A binding that uses a message part, task-state extension, or other carrier MUST preserve the same refusal semantics and scope the challenge to the identified attempt. It MUST NOT represent the challenge as a successful task result, output artifact, authorization, or durable state that silently applies to later attempts. These rules constrain carrier semantics without requiring every transport to use the same envelope.

HTTP Binding An HTTP origin server that refuses an action because authorization evidence is missing, stale, or unverifiable SHOULD return 403 Forbidden as defined by . The response body MUST be an application/problem+json Problem Details object as defined by . Status 428 is not used: it is defined for requests that the origin server requires to be conditional, not for authorization-evidence negotiation. If temporary overload is the sole reason for refusal, the origin server MUST NOT fabricate an authorization-evidence challenge. It uses 503 Service Unavailable, or another response appropriate to the actual failure, and MAY use Retry-After as defined by . A 403 AE-CHALLENGE is appropriate when evidence is actually insufficient, including when load pressure also makes pacing the corrected presentation necessary. The Problem Details type and title MUST have the values shown above. The HTTP status code and the optional status member MUST both be 403. The core challenge MUST appear in the evidence_challenge extension member. Human-readable detail text is advisory and MUST NOT be parsed for protocol semantics. The origin server MUST send Cache-Control: no-store. A generic intermediary can still transform an HTTP response, so a client MUST validate the Problem Details object and the authenticated origin before acting on hints or presenting evidence. When the core challenge contains retry_timing, the origin server SHOULD send a Retry-After field as defined by . Its HTTP-date or delay-seconds value MUST NOT permit a follow-up request earlier than not_before. If the header and core member identify different lower bounds, a client MUST use the later bound and then apply jitter_sec. Retry-After does not carry the jitter value and does not change the challenge expiry, authorize the action, or make a retry safe after an uncertain effect.

Informative DMSC Gateway Profile Agent gateways can carry the transport-neutral challenge during federation or handoff. The receiving gateway is the relying party: it derives the exact action, applies its own trust anchors and policy, and issues a challenge when evidence is missing, stale, or unverifiable. The sending gateway or agent can obtain evidence and return a new presentation using a DMSC-defined extension point or error carrier. If a gateway rewrites or translates the proposed action, the next gateway re-evaluates that locally derived action and, when necessary, issues its own binding. The prior gateway's action digest does not survive the rewrite as the receiving gateway's binding. Any claim that the two actions are equivalent requires a separately defined mapping or transformation proof. When a receiving gateway refuses an action because required authorization evidence is missing, stale, or unverifiable, it may return a structured evidence challenge identifying the exact action, outstanding evidence requirements, applicable freshness or status constraints, and supported presentation profiles. The challenge does not authorize the action or transfer admission ownership; conserved admission across gateway boundaries remains the separate requirement described in Section 7.8 of the proposed next revision of . In particular, AE-CHALLENGE does not prevent Gateway A and Gateway B from admitting the same single-use right. It does not copy, fence, or relinquish consumption state and does not establish which gateway owns admission during a partition. The separate handoff property is not a race to decide which gateway admitted first; it is an invariant that committed capacity is not over-allocated under any interleaving of admissions and releases. A DMSC handoff profile MUST solve that property separately. If the receiving gateway cannot establish the required exclusivity, its safe result is refusal or indeterminate, not acceptance based on the challenge.

Gateway Conformance Case Gateway A forwards a proposed action to Gateway B. Gateway B derives the exact action under its pinned action profile and refuses because a current approval artifact is missing. B returns an AE-CHALLENGE bound to B's action digest and local evidence requirement. A obtains and presents the artifact. B verifies it under B's trust anchors, matches it to the same action, and reevaluates B's policy. The case passes only if A never treats the challenge as authorization and B refuses an action-digest mismatch, stale status, replayed nonce, or unsupported presentation profile. A second case attempts concurrent admission of one single-use right at both gateways. AE-CHALLENGE alone MUST NOT be reported as passing that case. The case passes only when a separate conserved-admission mechanism establishes exclusivity, or when the receiving gateway refuses or returns indeterminate because exclusivity cannot be established.

Security Considerations Challenge forgery and evidence exfiltration. A forged challenge cannot authorize an action, but it can redirect a presenter, induce unnecessary approval work, or solicit sensitive evidence. Presenters MUST authenticate the relying party and audience before following obtain hints or disclosing evidence. A bare JSON object copied outside its authenticated carrier has no origin authenticity. Action substitution. The relying party computes the challenge action digest from the action it would execute and recomputes that binding before admission. A digest supplied by the presenter or an action profile selected by the presenter defeats the purpose of the protocol. Replay and denial of service. Durable body-bound nonce registration is required before challenge exposure. Consumption before evidence evaluation prevents concurrent replays, but an attacker who can submit a structurally valid response can consume a challenge. Deployments therefore need authenticated presenters, rate limits, short expiries, and a cheap path to issue a replacement challenge. Retry synchronization and load amplification. A precise challenge tells a presenter how to correct a request, so a population receiving similar refusals can retry in a synchronized burst. An issuer under pressure SHOULD combine recipient-specific retry timing, a non-zero jitter interval, admission control, and rate limiting. A presenter SHOULD retain exponential backoff or equivalent local load control even when a challenge supplies a lower bound. Timing guidance paces only a new evidence presentation and MUST NOT be interpreted as evidence satisfaction or authorization. Information disclosure. Evidence requirements, policy identifiers, action profiles, and obtain hints can reveal sensitive policy structure. A relying party SHOULD disclose only the minimum information needed for remediation. The HTTP binding requires Cache-Control: no-store. Stale facts and uncertain effects. A satisfied challenge proves only that the relying party's evidence requirement was met at evaluation time. It does not prove that external facts remain true, that an action executed, or that an uncertain action is safe to retry. Status rechecking, execution custody, effect observation, and reconciliation remain separate controls. Gateway split brain. An authenticated challenge from a receiving gateway does not transfer a single-use authorization or its consumption ledger. Implementations MUST NOT claim cross-gateway double-admission prevention merely because both gateways implement this challenge.

IANA Considerations IANA is requested to add the following entry to the "HTTP Problem Types" registry established by . That registry uses the Specification Required policy defined by ; this request does not require IETF Review or Standards Action.

Type URI:

https://iana.org/assignments/http-problem-types#ae-required

Title:

Authorization Evidence Required

Recommended HTTP status code:

403

Reference:

This document.

This revision withdraws the earlier request for a new application/authorization-evidence-challenge+json media type. The HTTP binding reuses the registered application/problem+json media type. Non-HTTP bindings define their own carriage without changing the core data model.

Interoperability Questions for Review This revision resolves the transport layering, HTTP status and media type, gateway-ownership boundary, acquisition-protocol coupling, retry synchronization, and per-hop binding questions raised through revision -03. Review is requested on the following remaining choices:

• Which Agent2Agent and AgentProto error carriers can preserve the complete challenge, authenticated issuer, audience, and retry timing, and whether either protocol needs an additional attempt identifier.
• Whether a DMSC gateway profile should carry the core object directly or reference it by an authenticated, digest-bound URI.

Acknowledgments Thanks to Sumit P. Ahuja for identifying synchronized retry and load amplification as a consequence of precise challenge guidance, and for suggesting retry timing and jitter. Thanks to Guigui Wang for implementation feedback on structured error payloads, per-hop action rebinding, and the separation of capability discovery from local authority. Thanks to Henri Sirkkavaara for review of replay ordering, concurrent-presentation behavior, diagnostic ordering, and portable identifier semantics.

Changes since -03

• Adds optional transport-neutral retry timing with an absolute lower bound and independently selected per-challenge jitter.
• Maps the lower bound to HTTP Retry-After while retaining jitter in the core challenge.
• Defines retry synchronization and load amplification as a denial of service concern without making a challenge an overload response.
• Recommends structured refusal or error payloads while preserving transport-neutral carrier semantics.
• Requires each delegation or rewriting hop to re-evaluate its local action and issue its own binding.
• Separates capability discovery from authority and local evidence sufficiency.
• Requires portable absolute-URI identifiers for evidence types, action profiles, and presentation profiles.
• Checks action agreement before expiry and nonce consumption, and defines a concurrent duplicate during evidence evaluation as a replay that cannot trigger a second evaluation or admission.
• States conserved admission as a committed-capacity invariant over every admission and release interleaving while leaving that mechanism outside AE-CHALLENGE.

Normative References Informative References

Relationship to Existing Mechanisms CAID is one mechanism for deriving and mapping exact material actions. AEB defines executor-side evidence verification, refusal, and one-time consumption. Bounded Capability Receipts define durable single-domain reservation and consumption. AEC is one possible presentation profile. None is required by the core challenge, and the challenge does not extend any of them across independently operated admission domains.

Implementation Status The EMILIA Protocol repository contains a same-team reference implementation of the AE-CHALLENGE-v1 core lifecycle, including relying-party-derived action binding, policy-derived requirements, durable body-bound registration, single-use consumption, expiry, follow-up challenges, and action-swap refusal. Revision -03 added a serializer and parser for the RFC 9457 HTTP binding and tests that pin status 403, application/problem+json, Cache-Control: no-store, and the evidence_challenge extension. Revision -04 specifies retry timing, per-hop rebinding, discovery separation, carrier semantics, portable absolute-URI identifiers, and action-before-consumption replay ordering. These additions are based in part on external implementer review, including experience with structured error payloads and per-verifier decision bindings. They are not yet implemented by the same-team reference implementation and are not claimed as an independent implementation of this specification. This is same-team implementation evidence. No independent implementation or interoperability is claimed. The implementation does not yet provide a DMSC carrier or conserved-admission mechanism. Its current local evidence and presentation identifiers are not the absolute URIs required by this revision.