Standards Interoperability Integration
Empirical findings, architectural synthesis, and test results from integrating the Kantara Initiative’s “UMA for Agents” (U4A) federated authorization profile with the CEQUS national identity and authority framework.
Federated autonomous systems cannot unilaterally expand authority granted by human principals. When a purchasing agent interacts with an external vendor, the effective permission is the non-delegable intersection of three independent boundary sets:
Neither the supplier’s authorization server (AS) nor the buyer’s principal authority can bypass the other. If the supplier AS issues an RPT, but the buyer’s delegation grant is revoked or the catalog price exceeds the buyer’s per-transaction limit, the transaction fails closed under the SQLite write lock before any tool effect commits.
1. Scope and Kantara U4A Profile Baseline
The Kantara Initiative’s User-Managed Access (UMA 2.0) work group, led by Eve Maler and Nick Gamb, published the candidate profile “UMA for Agents” (U4A) to govern federated machine-to-machine authorization across autonomous boundaries. In November 2026, the CEQUS engineering team implemented and evaluated the complete U4A profile against the CEQUS authority architecture across 15 test suites comprising 106 baseline tests and 31 advanced extension tests.
All external transactions in this integration are strictly synthetic: bounded to ≤ USD 100 within a synthetic maple syrup procurement catalog. Zero live funds, external credit facilities, or production directories are mutated.
2. UMA/U4A Federated Interoperability Architecture
The diagram below illustrates the end-to-end federated control path between the buyer domain (governed by C=US / CEQUS authority) and the supplier domain (governed by the Kantara U4A vendor AS and Resource Server):
3. Key Findings Across the Integration Dimensions
Dimension 1: Authorization Server Architecture & Perimeter Defense
The initial capability gate was evaluated against a stock enterprise Keycloak 26.8.2 server deployed in a WSL2 Docker container. The investigation surfaced critical architectural trade-offs:
- Introspection Binding Gaps in Stock AS: Stock Keycloak 26.8.2 issues RFC 8705 certificate-bound access tokens (
cnf["x5t#S256"]) during initial PKCE token issuance. However, standard RFC 7662 token introspection (/token/introspect) drops thecnfclaim unless custom Java Service Provider Interfaces (SPIs) are built and deployed into the Keycloak runtime. - Cross-Agent Token Theft at AS Perimeter: Under stock Keycloak, an attacker agent (Agent B) presenting a stolen initial token minted for Agent A to the UMA ticket exchange endpoint (
grant_type: urn:ietf:params:oauth:grant-type:uma-ticket) is issued an RPT bound to Agent B’s certificate! Keycloak does not compare the client TLS certificate against the initial bearer token’scnfclaim during ticket exchange. - The Solution — Native Lightweight Node.js AS (
security/uma/local/node-as/): To resolve these gaps without java container overhead, we engineered a native, lightweight Node.js Authorization Server:- Sub-Second Cold Start: Boots in < 280 ms (versus 25+ seconds for Keycloak container initialization).
- Native Introspection
cnf: Directly preserves and returnscnf["x5t#S256"]on RFC 7662 introspection responses. - Perimeter Defense: Compares the TLS peer certificate thumbprint against the initial token’s
cnfbinding at the AS boundary, returning HTTP 400invalid_grantbefore ticket exchange or RPT issuance can occur.
Dimension 2: Proof-of-Possession (PoP) Dual Support
CEQUS Resource Servers enforce strict proof-of-possession across two complementary IETF standards:
- RFC 8705 (mTLS Token Binding): Cryptographically binds the RPT to the client’s X.509 certificate SHA-256 thumbprint (
cnf["x5t#S256"]). Validated during the TLS 1.3 handshake. - RFC 9421 (HTTP Message Signatures): Per-request cryptographic signatures over
Signature,Signature-Input, andSignature-Agentheaders using Ed25519 or ECDSA keys, providing end-to-end WIMSE / AIMS compatibility when requests traverse intermediate proxies. - Fail-Closed Invariant: The receiver accepts either binding method and records
token_binding.methodin the causal Merkle audit log. Requests lacking both bindings fail closed immediately (missing_token_binding).
Dimension 3: IEEE 7012 Machine-Readable Bilateral Contracts
Kantara U4A anticipates machine-readable terms negotiation prior to granting access. Rather than relying on simple URL acknowledgment strings, we engineered a full bilateral contract engine (security/uma/terms.mjs):
- Supplier Term Proffer: The supplier AS proffers machine-readable clauses:
purpose(e.g.commercial_procurement),data_retention_days(maximum telemetry retention),audit_logging(mandatory_merkle),liability_cap_usd, andjurisdiction(e.g.US-VT,CA-QC). - Buyer Negotiation Engine: The purchasing agent evaluates proffered clauses against its principal’s local delegation policy. If data retention exceeds the buyer’s ceiling or jurisdiction is unapproved, the agent rejects the terms (
terms_policy_violation). If compliant, the agent countersigns a canonical bilateral acceptance artifact with its private key. - Enforcement: The AS challenges unaccepted ticket exchanges with HTTP 403
need_info. The accepted bilateral contract is recorded in causal evidence underu4a_terms_acceptance.method: "ieee_7012_negotiated".
Dimension 4: RFC 8693 Multi-Hop Agent Sub-Delegation
Complex enterprise purchasing requires primary procurement agents to delegate tasks to specialized sub-agents (e.g. logistics or inventory verifiers). We implemented RFC 8693 OAuth 2.0 Token Exchange for agent hierarchies (security/uma/token-exchange.mjs):
- Strict Scope Narrowing: Sub-agents request a narrowed subset of permissions (e.g.
order:viewfrom a parent holdingpurchase,order:view). Scope expansion is rejected at the AS with HTTP 400invalid_scope. - Audit Lineage (
actClaim): The child RPT preserves the root human buyer’ssubwhile embedding an RFC 8693act(actor) claim recording the delegating parent agent’s client ID. - Cascading Revocation: Revoking the parent RPT immediately invalidates all downstream sub-delegated child tokens.
Dimension 5: OpenID CAEP / SSE Backchannel Push-Revocation
Under conventional OAuth/UMA architectures, Resource Servers rely on cached token introspection or short TTLs (e.g. 1,000 ms), leaving a window of vulnerability between administrative revocation and local enforcement. We implemented OpenID Continuous Access Evaluation Profile (CAEP) push-revocation (security/uma/caep-receiver.mjs):
- mTLS Event Stream: Resource Servers maintain a persistent Server-Sent Events (SSE) connection to
/backchannel/sse. - RFC 8935 Security Event Tokens (SETs): When an owner revokes a token (
POST /owner/revoke-token), the AS broadcasts a signed JWT conforming tohttps://schemas.openid.net/secevent/caep/event-type/token-revokedcarrying the SHA-256 hash of the target token. - Zero-Latency Core Check: The RS maintains an in-memory hash set filter. In
security/delegation/core.mjs,checkResourceAuthorizationconsults this filter under the SQLite write lock and rejects revoked requests immediately withresource_authorization_revoked_by_caep, collapsing the 1,000 ms TTL window to zero latency.
Dimension 6: Live Owner Approval Queue & Supervisor Dashboard
When an agent requests operations exceeding automated policy thresholds (e.g. high-value orders or novel vendors), the ticket enters an asynchronous human approval state:
- Asynchronous State Machine: The AS transitions the ticket to
request_submittedand responds to the agent with HTTP 403request_submitted. - Bounded Exponential Polling: The agent polling client (
pollTicketinsecurity/uma/async-queue.mjs) polls with bounded backoff (1s → 2s → 4s, max 5 retries). - Real-Time Webhook & Dashboard: The AS dispatches a
ticket_pendingwebhook and surfaces the request on a live, dark-mode supervisor dashboard (GET /owner/dashboard). When the human supervisor clicks Approve or Deny (POST /owner/decide), the agent’s next poll immediately resolves with the minted RPT or fails closed withgrant_denied_by_owner.
4. Architectural Refinements Discovered During Empirical Testing
Deploying and verifying the complete UMA/U4A test matrix surfaced five key architectural discoveries that required modifying baseline assumptions:
4.1 AS Perimeter Defense (Eliminating Lazy Introspection)
Conventional UMA architectures assume the Authorization Server is identity-agnostic during ticket exchange, leaving token verification to the downstream Resource Server’s introspection call. When evaluated against stock Keycloak 26.8.2, an attacker agent (Agent B) presenting an initial token stolen from Agent A at /ticket/exchange received an RPT bound to Agent B’s certificate. Our lightweight Node AS eliminates this vulnerability by enforcing mutual TLS verification directly at the AS perimeter: the caller’s TLS certificate SHA-256 thumbprint must match the cnf["x5t#S256"] claim in the incoming subject_token, failing closed with HTTP 400 invalid_grant before ticket evaluation occurs.
4.2 Zero-Latency CAEP Push Revocation (Eliminating the 1,000ms Exposure Window)
Early drafts relied on short token TTLs (1,000 ms) or periodic introspection polling. In multi-worker race condition tests (revocation-ordering.test.mjs and caep-revocation.test.mjs), a concurrent process could commit a purchase during the 1,000 ms window after an administrative kill switch signal was triggered. The architecture was revised to use OpenID Continuous Access Evaluation Profile (CAEP) Security Event Tokens (RFC 8935 SETs) pushed over persistent Server-Sent Events (/backchannel/sse). The Resource Server and the CEQUS delegation core maintain an in-memory hash set filter that invalidates tokens in O(1) time under the SQLite transaction write lock, reducing revocation latency from 1,000 ms to sub-millisecond local enforcement.
4.3 Interactive Bilateral Contract State Machine (IEEE 7012 need_info)
Traditional OAuth architectures treat Terms of Service as unilateral click-wrap links. Autonomous AI agents cannot legally or deterministically consent to unconstrained vendor terms. In our implementation (security/uma/terms.mjs), the UMA HTTP 403 need_info challenge was formalized into an interactive bilateral contract negotiation: the supplier AS proffers machine-readable clauses (data retention ceilings, dispute jurisdiction, synthetic liability caps, and mandatory Merkle logging); the buyer agent evaluates the proffer against its principal’s delegation policy and submits a cryptographically signed bilateral countersignature before the AS mints the RPT.
4.4 Bounded Exponential Backoff for Human-in-the-Loop Supervision
When an agent requests permissions exceeding autonomous policy thresholds (e.g., first-time suppliers or order values approaching the $100 ceiling), synchronous HTTP connection holds cause worker starvation and gateway timeouts. We refined the protocol into an asynchronous supervisor queue: the AS returns HTTP 403 request_submitted and dispatches a webhook to the resource owner. The buyer agent enters a bounded exponential backoff polling loop (1s → 2s → 4s, maximum 5 attempts). When the human supervisor approves or denies the ticket via the live dashboard (POST /owner/decide), the agent’s subsequent poll immediately resolves or fails closed with grant_denied_by_owner.
4.5 Strict 1:1 Identity Mapping Invariant
Multi-tenant agent gateways frequently suffer from confused deputy attacks if certificate-to-client mappings permit many-to-one or one-to-many associations. In security/uma/fixtures.mjs and security/delegation/core.mjs, the mapping schema was tightened to enforce strict 1:1 bidirectional uniqueness: exactly one client ID per X.509 certificate thumbprint and exactly one certificate per client ID. Any unmapped client or lookalike certificate is rejected immediately with resource_authorization_identity_mismatch.
5. Standards Interoperability Matrix
| Standard / Specification | Governing Body | Architecture Role | Status | Implementation & Evidence |
|---|---|---|---|---|
| UMA 2.0 (Grant & Federated Authz) | Kantara Initiative | Federated ticket-to-token exchange across autonomous domains | TESTED | Tested on local Keycloak 26.8.2 and Node reference AS; 15 suites (112 tests) in security/uma/tests/TEST-PLAN.md. Closes token-theft gap via receiver-side certificate binding checks. |
| UMA for Agents (U4A Candidate) | Kantara Initiative | Agent-specific identity binding, async approvals, terms negotiation | UNIT-TESTED | Project reference implementation in security/uma/local/node-as/; dual PoP, IEEE 7012 contract engine, async queue, and Merkle evidence witness in testbed. Draft profile; not an enterprise product. |
| RFC 8705 (mTLS Token Binding) | IETF OAuth WG | Certificate-bound access tokens via cnf["x5t#S256"] |
TESTED | Validated against Keycloak 26 and native Node AS; verified across DER/PEM representations. |
| RFC 9421 (HTTP Message Signatures) | IETF HTTPbis WG | Per-request cryptographic request signatures over message headers | TESTED | security/uma/http-sig.mjs; Ed25519 signatures over Signature, Signature-Input, Signature-Agent. |
| IEEE 7012 (Machine-Readable Terms) | IEEE P7012 WG | Machine-negotiable bilateral procurement contracts | TESTED | security/uma/terms.mjs; automated proffer, negotiation, and bilateral countersigning verified in ieee-7012-engine.test.mjs. |
| RFC 8693 (OAuth 2.0 Token Exchange) | IETF OAuth WG | Multi-hop agent sub-delegation with scope narrowing and act claim |
TESTED | security/uma/token-exchange.mjs; hierarchical agent token exchange verified in subdelegation.test.mjs. |
| OpenID CAEP / SSE (RFC 8935 SETs) | OpenID Foundation | Backchannel push-revocation over mTLS Server-Sent Events | TESTED | security/uma/caep-receiver.mjs; zero-latency fail-closed revocation verified in caep-revocation.test.mjs. |
| RFC 7662 (Token Introspection) | IETF OAuth WG | Token active status and certificate binding verification | TESTED | Native Node AS returns active: true and cnf; verified in node-as.test.mjs. |
| W3C WebAuthn / FIDO2 | W3C / FIDO Alliance | Hardware-attested human supervisor approval and kill switch | TESTED | YubiKey hardware attestation, PIN-protected enrollment, Windows Hello passkey stop drills. |
| X.500 / LDAPv3 (RFC 4511) | ITU-T / IETF | Hierarchical registry under c=US and c=CA |
TESTED | OpenLDAP (slapd 2.6) live directory; chained cross-border queries between c=US and c=CA. |
6. Empirical Verification & Test Matrix
All automated test suites pass with a 100% success rate on clean development environments and GitHub Actions CI:
| Suite Command | Scope / Focus | Tests | Passed | Skipped | Duration |
|---|---|---|---|---|---|
npm run test:uma:node-as | Node AS, IEEE 7012, RFC 8693, CAEP, Owner Queue | 42 | 42 | 0 | 3.3s |
npm run test:uma:plan | 15-Suite Baseline Kantara U4A Test Plan | 112 | 112 | 0 | 51.2s |
npm run test:uma | UMA Core Contract & Purchase Service | 31 | 31 | 0 | 3.7s |
npm run test:vda:remote | Remote Execution & Fault Invariants | 60 | 58 | 2 (POSIX) | 8.3s |
npm run test:residue | Package Mirror & Filesystem Isolation | 7 | 7 | 0 | 9.4s |
npm run test:compliance | NIST / CMMC Requirement Traceability | 11 | 11 | 0 | 1.6s |
npm run test:delegation | Full CEQUS Delegation Engine | 594 | 588 | 6 | 39.8s |
7. Limits and Academic Classification
- Academic Research: This work was conducted as part of the Penn State AI-285 course project on national person and AI-agent registries. It represents a proposed conceptual architecture and is not certified for production deployment.
- Synthetic Boundaries: All purchase transactions are executed against a synthetic catalog with a USD 100 transaction ceiling; no real currency, banking networks, or credit lines are involved.
- Attribution: Engineering artifacts were authored using Antigravity (Google DeepMind) and Claude Code (Anthropic) in compliance with project governance guidelines in
AGENTS.md.
Sources & Technical References
- Kantara Initiative: User-Managed Access (UMA) 2.0 Grant for OAuth 2.0 Authorization; E. Maler & N. Gamb, UMA for Agents (U4A) Profile Draft.
- IETF Specifications: RFC 8705 (Mutual-TLS Profile for OAuth 2.0), RFC 9421 (HTTP Message Signatures), RFC 8693 (OAuth 2.0 Token Exchange), RFC 7662 (Token Introspection), RFC 8935 (Security Event Tokens over HTTP).
- IEEE Standards: IEEE P7012 (Standard for Machine-Readable Personal Privacy Terms).
- OpenID Foundation: Continuous Access Evaluation Profile (CAEP) / Shared Signals and Events (SSE).
- CEQUS Architecture: Understanding the Architecture, Enterprise Compatibility, Zero Trust Execution, Native-Hosted Agents; repository code under
security/uma/andsecurity/delegation/.