ietf-draft-analyzer/workspace/drafts/ietf-exec-audit/draft-nennemann-exec-audit.md

fullname: Christian Nennemann
organization: Independent Researcher
email: ietf@nennemann.de

normative: RFC7515: RFC7517: RFC7518: RFC7519: RFC9449: RFC9562: RFC9110:

informative: RFC6838: RFC8693: RFC8725: RFC9334: I-D.ietf-scitt-architecture: I-D.ietf-oauth-transaction-tokens: I-D.ietf-wimse-arch: I-D.ietf-wimse-s2s-protocol: OPENTELEMETRY: title: "OpenTelemetry Specification" target: https://opentelemetry.io/docs/specs/otel/ date: false author: - org: Cloud Native Computing Foundation SPIFFE: title: "SPIFFE ID" target: https://spiffe.io/docs/latest/spiffe-about/spiffe-concepts/ date: false SD-JWT: title: "Selective Disclosure for JWTs (SD-JWT)" target: https://datatracker.ietf.org/doc/draft-ietf-oauth-selective-disclosure-jwt/ seriesinfo: Internet-Draft: draft-ietf-oauth-selective-disclosure-jwt date: 2024 CERTIFICATE-TRANSPARENCY: title: "Certificate Transparency" target: https://datatracker.ietf.org/doc/html/rfc9162 seriesinfo: RFC: 9162 date: 2021 EU-AI-ACT: title: "Regulation (EU) 2024/1689 (AI Act)" target: https://eur-lex.europa.eu/eli/reg/2024/1689/oj date: 2024 author: - org: European Parliament and Council EU-DORA: title: "Regulation (EU) 2022/2554 (DORA)" target: https://eur-lex.europa.eu/eli/reg/2022/2554/oj date: 2022 author: - org: European Parliament and Council

--- abstract

This document defines Execution Audit Tokens (EATs), a JWT-based mechanism for recording task execution across distributed agent workflows that span multiple administrative domains, identity frameworks, and regulatory jurisdictions. Each EAT captures a single task, linked to predecessor tasks through a directed acyclic graph (DAG).

EATs are identity-framework agnostic and operate with any asymmetric key infrastructure including X.509 certificates, OAuth-based credentials, SPIFFE/WIMSE identities, DID-based systems, or plain JWK sets.

EATs support three assurance levels — Level 1 (unsigned JSON), Level 2 (JOSE asymmetric signing), and Level 3 (JOSE signing with audit ledger) — allowing deployments to choose the appropriate trade-off between simplicity and regulatory compliance.

For cross-domain workflows, EATs introduce regulatory profile metadata, federated audit ledger anchoring, selective disclosure at domain boundaries, and machine-readable audit evidence profiles to satisfy divergent compliance requirements across jurisdictions simultaneously.

--- middle

Introduction

Distributed agent workflows increasingly span multiple administrative domains, identity ecosystems, and regulatory jurisdictions. An investment bank's agent in Frankfurt may coordinate with a credit rating agency's agent in New York; a healthcare agent processing patient data under HIPAA may hand off results to a research agent operating under GDPR. In each case, the agents must maintain auditable execution records that satisfy the compliance requirements of every jurisdiction through which the workflow passes.

Existing identity frameworks authenticate agents and propagate authorization context, but they do not record what those agents actually did. Distributed tracing systems (e.g., OpenTelemetry) provide observability but lack cryptographic integrity, are controlled by the platform operator, and do not satisfy regulatory audit requirements. Authorization tokens (e.g., OAuth Transaction Tokens) propagate delegation chains but cannot represent the branching and convergence patterns of real workflows.

This document defines Execution Audit Tokens (EATs), a JWT-based mechanism that fills the gap between agent identity and execution accountability. Each EAT captures a single task, linked to predecessor tasks through a directed acyclic graph (DAG). EATs are designed to operate with any identity framework that provides asymmetric key pairs with verifiable binding to agent identity.

The key design goals are:

• Identity-framework independence: EATs work with X.509, OAuth, SPIFFE/WIMSE, DID, or plain JWK sets
• Cross-domain interoperability: A single DAG can span agents using different identity frameworks
• Regulatory flexibility: Machine-readable regulatory profiles allow different segments of a workflow to satisfy different compliance requirements
• Federated audit: Each domain operates its own audit ledger; cross-ledger anchoring enables end-to-end verification
• Privacy at boundaries: Selective disclosure allows verification of DAG structure without exposing regulated data across jurisdictions

Scope and Applicability

This document defines:

• The Execution Audit Token (EAT) format ({{eat-format}})
• Three assurance levels for EAT production and verification ({{level-1}}, {{level-2}}, {{level-3}})
• DAG structure for task dependency ordering ({{dag-validation}})
• An HTTP header for EAT transport ({{http-header}})
• Cross-domain audit model including regulatory profiles, federated ledger anchoring, and selective disclosure ({{cross-domain}})
• Audit ledger interface requirements ({{ledger-interface}})

The following are out of scope and are handled by the deployment's identity framework:

• Agent authentication and identity provisioning
• Key distribution and management
• Trust domain establishment and management
• Credential lifecycle management

Assurance Levels and Applicability

EATs support three assurance levels that govern how tokens are produced, transported, and verified. Each level builds on the payload structure defined in {{eat-payload}}.

Level 1 (L1) — Unsigned JSON:

EATs are plain JSON objects with no cryptographic signature. Integrity depends on transport security (TLS). L1 is intended for internal deployments within a single trust domain where non-repudiation is not required, such as development environments or internal microservice meshes.

Level 2 (L2) — JOSE Asymmetric Signing:

EATs are signed as JWS Compact Serialization tokens using asymmetric keys bound to the agent's identity credential. L2 provides non-repudiation and tamper detection. This is the baseline assurance level for cross-organization and peer-to-peer deployments.

Level 3 (L3) — JOSE Signing with Audit Ledger:

L3 extends L2 by requiring that every EAT be recorded in an audit ledger with cryptographic commitment (hash chain, Merkle tree, or equivalent). L3 is intended for regulated environments that require hash-committed audit trails with tamper-evident history.

Assurance level selection is orthogonal to human-in-the-loop (HITL) policy: any level may be combined with HITL requirements. Level selection guidance is provided in {{level-selection}}.

Conventions and Definitions

{::boilerplate bcp14-tagged}

The following terms are used in this document:

Agent:

An autonomous workload that executes tasks within a workflow.

Task:

A discrete unit of agent work that consumes inputs and produces outputs.

Directed Acyclic Graph (DAG):

A graph structure representing task dependency ordering where edges are directed and no cycles exist.

Execution Audit Token (EAT):

A token defined by this specification that records task execution details. At Level 2 and Level 3, an EAT is a signed JSON Web Token {{RFC7519}}. At Level 1, an EAT is an unsigned JSON object with the same payload structure.

Identity Credential:

A credential that proves an agent's identity, such as an X.509 certificate, an OAuth client credential, a SPIFFE Workload Identity Token, a DID document, or a JWK with out-of-band binding. See {{identity-binding}}.

Assurance Level:

One of three tiers (L1, L2, L3) governing how an EAT is produced, transported, and verified. See {{assurance-levels}}.

Audit Ledger:

An append-only, immutable log of all EATs within a workflow or set of workflows, used for audit and verification.

Trust Domain:

An administrative boundary within which agents share a common identity authority.

Domain Boundary:

The point at which a workflow crosses from one trust domain to another, potentially involving different identity frameworks, regulatory regimes, or audit requirements.

Regulatory Profile:

A machine-readable identifier specifying the regulatory framework governing an EAT, determining required claims, retention periods, and audit properties. See {{regulatory-profiles}}.

Ledger Anchor:

A signed receipt from one domain's audit ledger that can be verified by another domain's ledger, enabling cross-domain audit chain verification. See {{federated-ledger}}.

Actor Type:

A classification of the entity that performed the task: "agent" (autonomous workload), "human" (human operator), or "system" (infrastructure component). See {{actor-type-claim}}.

Execution Audit Token Format

An Execution Audit Token records a single task in a distributed workflow. The EAT payload structure is common across all assurance levels; only the envelope and verification requirements differ.

EAT Payload Structure

The EAT payload is a JSON object containing the claims defined in this section.

Standard Claims

The following standard JWT claims {{RFC7519}} are defined for EATs:

iss:

OPTIONAL at L1; REQUIRED at L2 and L3. StringOrURI. A URI identifying the issuer of the EAT. The value MUST correspond to the agent's identity as asserted by its identity credential (see {{identity-binding}}). Deployments MAY use HTTPS URLs, URN:UUID identifiers, SPIFFE IDs, DID URIs, or other URI schemes appropriate to the identity framework in use. L1 deployments are encouraged to include "iss" for consistency but it is not required. See {{l2-verification}} and {{l3-verification}} for the L2/L3 verification requirements.

aud:

RECOMMENDED. StringOrURI or array of StringOrURI. The intended recipient(s) of the EAT. The "aud" claim SHOULD contain the identifiers of all entities that will verify the EAT. When an EAT is to be verified by both the next agent and the audit ledger independently, "aud" MUST be an array containing both identifiers. Each verifier checks that its own identity appears in "aud". The "aud" claim is REQUIRED for L2 and L3 deployments (see {{l2-verification}} and {{l3-verification}}).

iat:

REQUIRED. NumericDate. The time at which the EAT was issued.

exp:

REQUIRED. NumericDate. The expiration time of the EAT. Implementations SHOULD set this to 5 to 15 minutes after "iat".

jti:

REQUIRED. String. A unique identifier for both the EAT and the task it records, in UUID format {{RFC9562}}. The "jti" serves as both the token identifier (for replay detection) and the task identifier (for DAG parent references in "pred"). Receivers MUST reject EATs whose "jti" has already been seen within the expiration window. When "wid" is present, uniqueness is scoped to the workflow; when "wid" is absent, uniqueness MUST be enforced globally across the EAT store.

Execution Context Claims

The following claims are defined by this specification:

wid:

OPTIONAL. String. A workflow identifier that groups related EATs into a single workflow. When present, MUST be a UUID {{RFC9562}}.

exec_act:

REQUIRED. String. The action or task type identifier describing what the agent performed (e.g., "process_payment", "validate_safety"). This claim name avoids collision with the "act" (Actor) claim registered by {{RFC8693}}.

pred:

REQUIRED. Array of strings. Predecessor task identifiers representing DAG dependencies. Each element MUST be the "jti" value of a previously verified EAT. An empty array indicates a root task with no dependencies. A workflow MAY contain multiple root tasks. The "pred" claim is always required (rather than optional with absence meaning "root task") to simplify validation logic and eliminate ambiguity between a root task and a claim accidentally omitted.

Actor Type Claim

actor_type:

OPTIONAL. String. Identifies the type of entity that performed the task. Values are drawn from a controlled vocabulary:

• "agent": An autonomous software workload (default if omitted)
• "human": A human operator who performed a manual step, such as an approval or review
• "system": An infrastructure component acting automatically, such as a load balancer, gateway, or scheduler

When "actor_type" is "human", the "iss" claim SHOULD identify the human's organizational identity (e.g., an email-based URI or employee identifier URI) rather than a workload identity. Human-performed tasks are typically approval or review steps in human-in-the-loop (HITL) workflows.

At L2 and L3, EATs with "actor_type" of "human" MUST still be signed with an asymmetric key. The signing key MAY belong to the human's personal credential or to the system through which the human performed the action (e.g., an approval gateway), with the "iss" claim identifying the human and the signing system documented in "eat_ext".

Data Integrity Claims

The following claims provide integrity verification for task inputs and outputs without revealing the data itself:

inp_hash:

OPTIONAL. String. The base64url encoding (without padding) of the SHA-256 hash of the input data, computed over the raw octets of the input. SHA-256 is the mandatory algorithm with no algorithm prefix in the value, consistent with {{RFC9449}}.

out_hash:

OPTIONAL. String. The base64url encoding (without padding) of the SHA-256 hash of the output data, using the same format as "inp_hash".

Domain Metadata Claims

The following claims support cross-domain workflows:

domain:

OPTIONAL. String. The trust domain in which the agent operates. This is a DNS-compatible identifier (e.g., "bank.example.com", "hospital.example.org"). When present, enables auditors to determine which regulatory regime governs this segment of the DAG.

eat_reg:

OPTIONAL. Object. Regulatory profile metadata for this EAT. See {{regulatory-profiles}} for the structure and semantics.

ledger_receipt:

OPTIONAL. Object. A signed receipt from the domain's audit ledger, enabling cross-domain verification. See {{federated-ledger}} for the structure and semantics. This claim is populated by the ledger upon recording, not by the producing agent.

Extension Claims

eat_ext:

OPTIONAL. Object. A general-purpose extension object for domain-specific claims not defined by this specification. Implementations that do not understand extension claims MUST ignore them. Extension key names SHOULD use reverse domain notation (e.g., "com.example.custom_field") to avoid collisions. The serialized "eat_ext" object MUST NOT exceed 4096 bytes and MUST NOT exceed a nesting depth of 5 levels. Receivers MUST reject EATs whose "eat_ext" object exceeds these limits.

Complete EAT Payload Example

The following is a complete EAT payload example showing a cross-domain task with regulatory metadata. This payload structure is the same at all assurance levels; only the transport envelope differs.

{
  "iss": "https://agents.bank.example.com/compliance-agent",
  "aud": ["https://agents.partner.example.com/rating-agent",
          "https://ledger.bank.example.com"],
  "iat": 1772064150,
  "exp": 1772064750,
  "jti": "550e8400-e29b-41d4-a716-446655440001",

  "wid": "a0b1c2d3-e4f5-6789-abcd-ef0123456789",
  "exec_act": "verify_trade_compliance",
  "pred": ["550e8400-e29b-41d4-a716-446655440000"],
  "actor_type": "agent",

  "domain": "bank.example.com",
  "eat_reg": {
    "profiles": ["urn:ietf:eat:reg:eu-dora",
                  "urn:ietf:eat:reg:eu-ai-act"],
    "retention_days": 2555,
    "jurisdiction": "EU"
  },

  "inp_hash": "n4bQgYhMfWWaL-qgxVrQFaO_TxsrC4Is0V1sFbDwCgg",
  "out_hash": "LCa0a2j_xo_5m0U8HTBBNBNCLXBkg7-g-YpeiGJm564",

  "eat_ext": {
    "com.example.trace_id": "abc123"
  }
}

{: #fig-full-eat title="Complete EAT Payload Example"}

Level 1: Unsigned JSON

At Level 1, an EAT is a plain JSON object carrying the payload defined in {{eat-payload}}. No cryptographic signature is applied.

L1 Transport

L1 EATs are transported as base64url-encoded JSON (without padding). Two mechanisms are defined:

HTTP Header:

The Execution-Context header field ({{http-header}}) carries the base64url-encoded JSON payload (without padding). The header value is NOT raw JSON; the receiver MUST base64url-decode the value before parsing.

HTTP Body:

When the EAT is the primary request payload, the EAT MAY be carried in the HTTP request body with Content-Type application/json.

L1 EATs MUST be transmitted over TLS or mTLS connections. Transport security is the sole integrity mechanism at this level.

L1 Verification

When an agent receives an L1 EAT, it MUST perform the following verification steps:

1. Parse the JSON object.

2. Verify that all required claims ("jti", "iat", "exp", "exec_act", "pred") are present and well-formed.

3. Verify that the "jti" has not been previously seen within the expiration window, consistent with the replay detection requirement in {{exec-claims}}.

4. Verify the "exp" claim indicates the EAT has not expired.

5. Verify the "iat" claim is not unreasonably far in the past (RECOMMENDED maximum: 15 minutes) and is not unreasonably far in the future (RECOMMENDED: no more than 30 seconds ahead of the verifier's current time).

6. Perform DAG validation per {{dag-validation}}.

If any verification step fails, the EAT MUST be rejected and the failure MUST be logged.

L1 Security Properties and Applicability

L1 provides the following security properties:

• Structural integrity via JSON schema validation
• Temporal ordering via "iat" and "exp" claims
• DAG integrity via parent reference validation
• Transport integrity via TLS

L1 does NOT provide:

• Non-repudiation (no signature binds the EAT to its issuer)
• Tamper detection after delivery (the recipient cannot prove the EAT was not modified after receipt)
• Issuer authentication at the EAT layer (identity depends entirely on the transport layer)

Note: At L1, without a deployed audit ledger, DAG parent existence validation ({{dag-validation}} step 2) is limited to parent EATs received inline with the current request. Parents from prior requests that were not forwarded inline cannot be verified.

L1 MUST NOT be used across trust domain boundaries. Deployments using L1 SHOULD restrict it to internal environments where all agents are operated by the same organization and transport security is considered sufficient.

Level 2: JOSE Asymmetric Signing

At Level 2, an EAT is a JSON Web Token (JWT) {{RFC7519}} signed as a JSON Web Signature (JWS) {{RFC7515}}.

EATs MUST use JWS Compact Serialization (the base64url-encoded header.payload.signature format) so that they can be carried in a single HTTP header value.

The EAT MUST be signed with a private key that is verifiably bound to the agent's identity credential (see {{identity-binding}}). The JOSE header "kid" parameter MUST reference the public key identifier from the agent's identity credential, and the "alg" parameter MUST match the algorithm associated with that credential.

L2 JOSE Header

The EAT JOSE header MUST contain the following parameters:

{
  "alg": "ES256",
  "typ": "eat+jwt",
  "kid": "agent-a-key-id-123"
}

{: #fig-header title="EAT JOSE Header Example"}

alg:

REQUIRED. The digital signature algorithm used to sign the EAT. MUST match the algorithm associated with the agent's identity credential. Implementations MUST support ES256 {{RFC7518}}. The "alg" value MUST NOT be "none". Symmetric algorithms (e.g., HS256, HS384, HS512) MUST NOT be used, as EATs require asymmetric signatures for non-repudiation. To support algorithm agility, deployments SHOULD maintain an allowlist of accepted signing algorithms and SHOULD plan for migration to stronger algorithms as cryptographic requirements evolve.

typ:

REQUIRED. MUST be set to "eat+jwt" to distinguish EATs from other JWT types.

kid:

REQUIRED. The key identifier referencing the public key from the agent's identity credential {{RFC7517}}. Used by verifiers to look up the correct public key for signature verification.

L2 Transport

L2 EATs are transported in the Execution-Context HTTP header field ({{http-header}}) as JWS Compact Serialization. The header field value consists of three Base64url-encoded parts separated by period (".") characters.

L2 EATs MUST be transmitted over TLS or mTLS connections.

L2 Verification

When an agent receives an L2 EAT, it MUST perform the following verification steps in order:

1. Parse the JWS Compact Serialization to extract the JOSE header, payload, and signature components per {{RFC7515}}.

2. Verify that the "typ" header parameter is "eat+jwt".

3. Verify that the "alg" header parameter appears in the verifier's configured allowlist of accepted signing algorithms. The allowlist MUST NOT include "none" or any symmetric algorithm (e.g., HS256, HS384, HS512). Implementations MUST include ES256 in the allowlist; additional asymmetric algorithms MAY be included per deployment policy.

4. Verify the "kid" header parameter references a known, valid public key from an identity credential within the trust domain or a federated trust domain (see {{identity-binding}}).

5. Retrieve the public key identified by "kid" and verify the JWS signature per {{RFC7515}} Section 5.2.

6. Verify that the signing key identified by "kid" has not been revoked within the trust domain. Implementations MUST check the key's revocation status using the identity framework's key lifecycle mechanism (e.g., certificate revocation list, OCSP, SPIFFE trust bundle updates, JWK set rotation, or DID document deactivation).

7. Verify the "alg" header parameter matches the algorithm associated with the agent's identity credential.

8. Verify the "iss" claim is present and matches the agent identity asserted by the identity credential associated with the "kid" public key (see {{identity-binding}}).

9. Verify the "aud" claim is present and contains the verifier's own identity. When "aud" is an array, it is sufficient that the verifier's identity appears as one element; the presence of other audience values does not cause verification failure. When the verifier is the audit ledger, the ledger's own identity MUST appear in "aud".

10. Verify the "exp" claim indicates the EAT has not expired.

11. Verify the "iat" claim is not unreasonably far in the past (implementation-specific threshold, RECOMMENDED maximum of 15 minutes) and is not unreasonably far in the future (RECOMMENDED: no more than 30 seconds ahead of the verifier's current time, to account for clock skew).

12. Verify all required claims ("jti", "exec_act", "pred") are present and well-formed.

13. If "actor_type" is present, verify it contains a recognized value ("agent", "human", or "system").

14. If "eat_reg" is present, verify the regulatory profile structure per {{regulatory-profiles}}.

15. Perform DAG validation per {{dag-validation}}.

16. If all checks pass and an audit ledger is deployed, the EAT SHOULD be appended to the ledger.

If any verification step fails, the EAT MUST be rejected and the failure MUST be logged for audit purposes. Error messages SHOULD NOT reveal whether specific parent task IDs exist in the EAT store, to prevent information disclosure.

L2 Security Properties and Applicability

L2 provides the following security properties:

• Non-repudiation (the EAT is cryptographically bound to the issuing agent via asymmetric signature)
• Tamper detection (any modification invalidates the signature)
• Issuer authentication (the "kid" links to the agent's identity credential)
• Audience restriction (the "aud" claim limits valid verifiers)
• All L1 properties (structural, temporal, DAG, transport integrity)

L2 is suitable for cross-organization deployments, peer-to-peer agent communication, and any deployment requiring cryptographic non-repudiation. L2 is the RECOMMENDED default for production deployments.

Level 3: JOSE Signing with Audit Ledger

Level 3 extends Level 2 by requiring that every EAT be recorded in an audit ledger that provides cryptographic commitment. L3 transport and JWS signing are identical to L2; the additional requirements apply to ledger recording and verification.

L3 Transport

L3 EATs use the same JWS Compact Serialization transport as L2 (see {{l2-transport}}).

L3 Audit Ledger Requirements

At Level 3, the audit ledger MUST satisfy the baseline requirements defined in {{ledger-interface}} and the following additional requirements:

1. Hash Chain: Each ledger entry MUST include a cryptographic hash of the previous entry, forming a hash chain. The hash algorithm MUST be SHA-256 or stronger.

2. Cryptographic Commitment: The ledger MUST provide a mechanism for verifiable cryptographic commitment over its entries, such as Merkle tree proofs, hash chains with signed roots, or equivalent constructions. The commitment MUST allow a verifier to confirm that a specific EAT is included in the ledger at a specific position without trusting the ledger operator.

3. Receipts: Upon successful append, the ledger MUST return a receipt to the submitting agent. The receipt MUST contain the ledger sequence number, the hash of the appended EAT entry, and the cryptographic commitment proof (e.g., Merkle inclusion proof) at the time of recording.

4. Availability: The ledger MUST be accessible for verification queries by authorized entities within the trust domain.

L3 Recording Semantics

An agent producing an L3 EAT MUST attempt to record the EAT in the audit ledger. Two recording modes are defined:

Synchronous Recording:

The agent submits the EAT to the ledger and waits for a receipt before transmitting the EAT to the next agent. Synchronous recording provides the strongest guarantee: the EAT is committed to the ledger before any downstream processing occurs.

Asynchronous Recording:

The agent submits the EAT to the ledger and transmits the EAT to the next agent without waiting for a receipt. The agent MUST subsequently verify that the ledger accepted the EAT and MUST retain the receipt. If the ledger rejects the EAT, the agent MUST alert the workflow coordinator or log a critical error. If asynchronous ledger recording fails, the producing agent MUST notify downstream agents that the EAT's L3 status is unconfirmed. Downstream agents SHOULD treat such an EAT as L2-verified (not L3-verified) until ledger confirmation is independently obtainable.

Deployments SHOULD use synchronous recording unless latency constraints make it impractical. The recording mode SHOULD be documented in the deployment's security policy.

L3 Verification

L3 verification consists of L2 verification ({{l2-verification}}) followed by ledger verification:

Note: The "iss" and "aud" claims, which are RECOMMENDED at L1 and verified at L2 (steps 8-9 of {{l2-verification}}), are REQUIRED at L3.

1. Perform all L2 verification steps (steps 1 through 16 of {{l2-verification}}).

2. Verify that the EAT has been recorded in the audit ledger by querying the ledger for the EAT's "jti". If the EAT is found, verify that the ledger entry's receipt contains a valid sequence number, EAT hash, and cryptographic commitment proof. Note: the producing agent is responsible for recording the EAT per {{l3-recording}}; the verifier checks that recording has occurred.

3. If the EAT is not yet present in the ledger (e.g., due to asynchronous recording), the verifier MAY retry after a short delay. If the ledger does not contain the EAT within the configured timeout, the verifier MUST either reject the EAT or downgrade to L2 verification per deployment policy.

4. If the ledger is unavailable, the verifier SHOULD retry with exponential backoff. If the ledger remains unavailable after a deployment-configured number of retries, the verifier MUST either reject the EAT or downgrade to L2 verification per deployment policy.

If any L3-specific verification step fails, the EAT MUST be rejected even if L2 verification succeeded.

L3 Security Properties and Applicability

L3 provides the following security properties beyond L2:

• Tamper-evident history (the hash chain detects any modification, insertion, or deletion of ledger entries)
• Cryptographic commitment (Merkle proofs or equivalent allow independent verification of ledger inclusion)
• Non-repudiation of recording (the receipt proves the EAT was committed at a specific time and position)
• All L2 properties

L3 is intended for regulated environments requiring auditable, tamper-evident execution records, such as healthcare (FDA 21 CFR Part 11), financial services (MiFID II, DORA {{EU-DORA}}), and environments subject to the EU AI Act {{EU-AI-ACT}}.

Level Selection

Deployments SHOULD select the assurance level based on the following criteria:

• Use L1 when all agents operate within a single trust domain, non-repudiation is not required, and transport security (TLS) is considered sufficient. Typical environments: development, testing, internal microservice meshes.

• Use L2 when agents cross trust domain boundaries, when non-repudiation is required, or when agents communicate peer-to-peer across organizations. L2 is the RECOMMENDED default for production deployments.

• Use L3 when regulatory requirements mandate tamper-evident audit trails with cryptographic commitment, or when the deployment needs to demonstrate compliance with regulatory frameworks.

A deployment MAY use different assurance levels for different workflows within the same infrastructure. When agents at different levels interact, the higher level's verification requirements apply to the receiving agent. Specifically, an EAT at a higher assurance level MAY reference parent EATs at a lower assurance level in its "pred" claim. In this case, the receiving agent applies its own level's verification to the current EAT and the parent's level verification to each parent EAT. Whether cross-level parent references are permitted is a deployment policy decision; deployments MAY reject EATs whose parents are below a minimum assurance level.

This specification does not define a level negotiation mechanism. Deployments configure the required assurance level out of band. A future extension MAY define in-band level signaling.

Assurance Level Transitions at Domain Boundaries

When a workflow crosses a domain boundary, the required assurance level may change. For example, an internal workflow segment may use L1 within an organization, but L2 is required when crossing to an external partner, and L3 is required when entering a regulated jurisdiction.

Deployments SHOULD document their assurance level policy for cross-domain interactions. When an EAT crosses a domain boundary and the receiving domain requires a higher assurance level than the sending domain:

• The receiving agent MUST produce its own EAT at the required level, referencing the sender's EAT in "pred".

• The receiving agent MAY verify the sender's EAT at the sender's assurance level, provided that level meets the receiver's minimum acceptable level for parent EATs.

• If the sender's assurance level is below the receiver's minimum for parent EATs, the receiver MUST reject the EAT.

Backward Compatibility and Level Detection

A verifier determines the assurance level of a received EAT as follows:

1. If the raw Execution-Context header value (or body content) contains exactly two period (".") characters separating three non-empty segments, attempt to parse the value as JWS Compact Serialization. If the first segment base64url-decodes to a JSON object containing an "alg" field, the EAT is L2 or L3.

2. Otherwise, base64url-decode the header value (without padding) and attempt to parse the result as JSON. If successful, the EAT is L1.

3. If neither parse succeeds, the EAT MUST be rejected.

L2 and L3 use the same JWS format. Differentiation between L2 and L3 is a matter of deployment policy: L3 deployments require ledger recording ({{l3-ledger}}), while L2 deployments treat ledger recording as optional.

Identity Binding

At L2 and L3, the EAT signing key MUST be verifiably bound to the agent's identity. This specification defines the abstract requirements for identity binding; concrete bindings are defined for specific identity frameworks.

An identity binding MUST satisfy the following requirements:

1. Key Association: The "kid" value in the EAT JOSE header MUST resolve to a public key through the identity framework's key discovery mechanism.

2. Issuer Correspondence: The "iss" claim in the EAT payload MUST correspond to the agent identity asserted by the identity credential containing the "kid" public key.

3. Algorithm Consistency: The "alg" value in the EAT JOSE header MUST match the algorithm associated with the identity credential.

4. Revocation Checking: The verifier MUST be able to determine whether the signing key has been revoked using the identity framework's revocation mechanism.

Cross-Domain Identity Resolution

When an EAT crosses a domain boundary, the receiving domain may use a different identity framework than the sending domain. For cross-domain verification:

1. The receiving domain MUST be able to obtain the sender's public key through a pre-established federation mechanism (e.g., shared JWK set endpoint, cross-certified X.509 hierarchy, bilateral trust bundle exchange).

2. The "iss" claim's URI scheme indicates the identity framework in use. The receiving domain MUST support verification of the identity framework indicated by the "iss" URI scheme, or MUST reject the EAT.

3. Deployments SHOULD document which identity frameworks are accepted at each domain boundary and how cross-domain key discovery is configured.

X.509 Binding

When the deployment uses X.509 certificates:

• The EAT "kid" MAY be the certificate thumbprint (x5t) or subject key identifier.
• The EAT "iss" SHOULD be a URI derived from the certificate subject or subject alternative name.
• Revocation is checked via CRL or OCSP.

OAuth Binding

When the deployment uses OAuth 2.0 credentials:

• The EAT "kid" MUST match the "kid" parameter in the authorization server's published JWK set.
• The EAT "iss" SHOULD be the OAuth client identifier as a URI.
• Revocation is checked via token introspection or JWK set rotation.

JWK Set Binding

When the deployment uses plain JWK sets {{RFC7517}} with out-of-band trust establishment:

• The EAT "kid" MUST match the "kid" parameter in a JWK from the trusted JWK set.
• The EAT "iss" MUST match a value associated with the JWK through the deployment's configuration.
• Revocation is handled by removing the key from the published JWK set.

DID Binding

When the deployment uses Decentralized Identifiers (DIDs):

• The EAT "kid" MUST reference a verification method in the DID document.
• The EAT "iss" MUST be the DID URI.
• Revocation is checked via DID document deactivation or verification method removal.

Cross-Domain Audit Model

This section defines mechanisms for maintaining verifiable audit trails across multiple administrative domains with potentially different regulatory requirements.

Regulatory Profiles

The "eat_reg" claim carries machine-readable metadata about the regulatory regime governing an EAT. This enables automated compliance checking: an auditor or receiving domain can determine what rules apply to a given segment of the DAG without out-of-band knowledge.

Regulatory Profile Structure

The "eat_reg" claim is a JSON object with the following fields:

profiles:

REQUIRED (within "eat_reg"). Array of strings. Each string is a URI identifying a regulatory profile. This specification defines an initial set of profile URIs in {{profile-registry}}; additional profiles may be registered per {{iana-profiles}}.

retention_days:

OPTIONAL. Integer. The minimum number of days this EAT and its associated audit records MUST be retained. When multiple regulatory profiles apply, the value SHOULD be the maximum retention period across all applicable profiles.

jurisdiction:

OPTIONAL. String. An ISO 3166-1 alpha-2 country code or regional identifier (e.g., "EU", "US") indicating the jurisdiction under which this EAT was produced.

redact_policy:

OPTIONAL. String. A URI identifying the data minimization or redaction policy applicable to this EAT when it crosses domain boundaries. See {{selective-disclosure}}.

Regulatory Profile Semantics

Each regulatory profile URI defines:

• Which EAT claims are REQUIRED vs OPTIONAL under that profile
• Whether "inp_hash" and "out_hash" are mandatory
• Whether L3 ledger recording is mandatory
• Minimum retention periods
• Whether selective disclosure is permitted or required at domain boundaries

Deployments MUST ensure that all EATs comply with every profile listed in their "eat_reg.profiles" array. When profiles conflict (e.g., one requires data retention while another requires data deletion), the deployment MUST resolve the conflict per its legal counsel and document the resolution.

Example Regulatory Profile Usage

An EAT produced in the EU financial sector:

{
  "eat_reg": {
    "profiles": [
      "urn:ietf:eat:reg:eu-dora",
      "urn:ietf:eat:reg:eu-ai-act"
    ],
    "retention_days": 2555,
    "jurisdiction": "EU"
  }
}

{: #fig-reg-eu title="EU Financial Regulatory Profile"}

An EAT produced in the US healthcare sector:

{
  "eat_reg": {
    "profiles": [
      "urn:ietf:eat:reg:us-hipaa"
    ],
    "retention_days": 2190,
    "jurisdiction": "US",
    "redact_policy": "urn:ietf:eat:redact:hipaa-minimum"
  }
}

{: #fig-reg-us title="US Healthcare Regulatory Profile"}

Federated Audit Ledger Model

In cross-domain workflows, each domain operates its own audit ledger. The federated ledger model enables end-to-end audit verification across domain boundaries without requiring a single shared ledger.

Cross-Ledger Anchoring

When an EAT crosses a domain boundary, the sending domain's ledger produces a signed anchor receipt that the receiving domain's ledger can reference. This creates a chain of cryptographic commitments across ledgers, analogous to how Certificate Transparency logs {{CERTIFICATE-TRANSPARENCY}} cross-reference through signed timestamps.

The anchor receipt is carried in the "ledger_receipt" claim of the EAT:

{
  "ledger_receipt": {
    "ledger_id": "https://ledger.bank.example.com",
    "seq": 42,
    "eat_hash": "n4bQgYhMfWWaL-qgxVrQFaO_TxsrC4Is0V1sFbDwCgg",
    "commitment": "LCa0a2j_xo_5m0U8HTBBNBNCLXBkg7-g-YpeiGJm564",
    "timestamp": 1772064200,
    "sig": "MEUCIQDx...signed-by-ledger-key..."
  }
}

{: #fig-anchor title="Cross-Ledger Anchor Receipt"}

ledger_id:

REQUIRED (within "ledger_receipt"). String. A URI identifying the audit ledger that recorded the EAT.

seq:

REQUIRED (within "ledger_receipt"). Integer. The sequence number of the EAT in the ledger.

eat_hash:

REQUIRED (within "ledger_receipt"). String. The base64url encoding (without padding) of the SHA-256 hash of the recorded EAT.

commitment:

REQUIRED (within "ledger_receipt"). String. The cryptographic commitment proof (e.g., Merkle root hash) at the time of recording.

timestamp:

REQUIRED (within "ledger_receipt"). NumericDate. The time at which the ledger recorded the EAT.

sig:

REQUIRED (within "ledger_receipt"). String. The base64url encoding of the ledger's signature over the concatenation of ledger_id, seq, eat_hash, commitment, and timestamp, using the ledger's signing key.

Cross-Ledger Verification

When a receiving domain encounters an EAT with a "ledger_receipt" from a foreign ledger, it MUST:

1. Verify the "sig" field using the foreign ledger's public key, obtained through a pre-established trust relationship.

2. Verify that the "eat_hash" matches the SHA-256 hash of the received EAT.

3. Optionally query the foreign ledger to verify the commitment proof.

4. Record the foreign EAT and its receipt in the local ledger, creating a cross-ledger reference.

Ledger Federation Requirements

For federated ledger operation:

1. Each domain's ledger MUST publish its signing public key at a well-known endpoint or through a pre-established key exchange mechanism.

2. Domains MUST establish bilateral or multilateral trust relationships specifying which foreign ledgers are accepted.

3. Ledger receipts MUST be verifiable without requiring real-time access to the foreign ledger (the receipt is self-contained for basic verification).

Selective Disclosure at Domain Boundaries

When an EAT crosses a domain boundary, the receiving domain may need to verify the DAG structure without seeing all details from the originating domain. This is critical for regulatory compliance: a HIPAA-governed EAT should not expose protected health information (PHI) to a receiving domain that operates under different regulations.

SD-JWT Integration

At L2 and L3, deployments MAY use Selective Disclosure JWT (SD-JWT) {{SD-JWT}} to enable selective disclosure of EAT claims at domain boundaries. When SD-JWT is used:

1. The producing agent creates an SD-JWT-based EAT where sensitive claims (e.g., "eat_ext", "inp_hash", "out_hash", "exec_act" details) are selectively disclosable.

2. At the domain boundary, the agent creates a presentation that discloses only the claims needed by the receiving domain:

   • DAG structure claims ("jti", "pred", "wid") are always disclosed
   • "iss" and "aud" are always disclosed for verification
   • "domain" and "eat_reg" are always disclosed for regulatory routing
   • Other claims are disclosed only as required by the receiving domain's policy

3. The receiving domain can verify the EAT's signature and DAG position without accessing redacted claims.

Minimum Disclosure Set

When selective disclosure is used, the following claims MUST always be disclosed (the "minimum disclosure set"):

• "jti" (task identity and DAG reference)
• "iat" and "exp" (temporal validation)
• "pred" (DAG structure)
• "iss" (issuer verification)
• "domain" (regulatory routing, when present)
• "eat_reg" (regulatory profile, when present)
• "actor_type" (when present)

Audit Evidence Profiles

An Audit Evidence Profile is a machine-readable specification of the EAT claims and verification properties required by a particular regulatory framework. Evidence profiles allow automated compliance checking: given an EAT, a compliance tool can determine whether it satisfies a specific regulatory requirement.

Evidence Profile Structure

An Audit Evidence Profile specifies:

• Profile URI (unique identifier)
• Required claims (which EAT claims must be present)
• Required assurance level (minimum L1, L2, or L3)
• Retention period (minimum days)
• Ledger requirements (whether L3 ledger is mandatory)
• Selective disclosure policy (whether disclosure is permitted at boundaries, and minimum disclosure set)
• Hash requirements (whether "inp_hash" and "out_hash" are mandatory)

Initial Evidence Profiles

This specification defines the following initial regulatory profile URIs. These profiles are informational and are intended to illustrate the framework; deployments MUST verify that these profiles meet their actual regulatory obligations.

urn:ietf:eat:reg:eu-dora:

EU Digital Operational Resilience Act. Minimum assurance level: L3. Retention: 2555 days (7 years). "inp_hash" and "out_hash": REQUIRED. Ledger: REQUIRED.

urn:ietf:eat:reg:eu-ai-act:

EU Artificial Intelligence Act. Minimum assurance level: L2. Retention: 3650 days (10 years) for high-risk AI systems. "inp_hash" and "out_hash": REQUIRED. Ledger: RECOMMENDED. "actor_type": REQUIRED.

urn:ietf:eat:reg:us-hipaa:

US Health Insurance Portability and Accountability Act. Minimum assurance level: L2. Retention: 2190 days (6 years). Selective disclosure at domain boundaries: REQUIRED for PHI-containing extensions.

urn:ietf:eat:reg:us-soc2:

SOC 2 Type II. Minimum assurance level: L2. Retention: 365 days (1 year).

urn:ietf:eat:reg:financial-mifid2:

MiFID II. Minimum assurance level: L3. Retention: 2555 days (7 years). Ledger: REQUIRED.

HTTP Header Transport

Execution-Context Header Field

This specification defines the Execution-Context HTTP header field {{RFC9110}} for transporting EATs between agents.

The format of the header field value depends on the assurance level:

• At Level 1, the header field value is the base64url-encoded JSON payload (without padding).

• At Level 2 and Level 3, the header field value is the EAT in JWS Compact Serialization format {{RFC7515}}. The value consists of three Base64url-encoded parts separated by period (".") characters.

An agent sending a request to another agent includes the Execution-Context header alongside any identity headers required by the deployment's identity framework.

GET /api/safety-check HTTP/1.1
Host: safety-agent.example.com
Authorization: Bearer <identity-token>
Execution-Context: eyJhbGci...EAT...

{: #fig-http-example title="HTTP Request with EAT Header"}

When multiple parent tasks contribute context to a single request, multiple Execution-Context header field lines MAY be included, each carrying a separate EAT. This applies to all assurance levels.

When a receiver processes multiple Execution-Context headers, it MUST individually verify each EAT per the applicable verification procedure ({{l1-verification}}, {{l2-verification}}, or {{l3-verification}}). If any single EAT fails verification, the receiver MUST reject the entire request. The set of verified parent task IDs across all received EATs represents the complete set of parent dependencies available for the receiving agent's subsequent EAT.

HTTP Error Handling

When EAT verification fails during HTTP request processing, the receiving agent SHOULD respond with HTTP 403 (Forbidden). This applies regardless of whether the failure is due to an invalid EAT payload, a signature verification failure, or a missing EAT when one is required by deployment policy. HTTP 401 (Unauthorized) SHOULD NOT be used for EAT failures, as 401 conventionally indicates that authentication credentials are missing or invalid and requires a WWW-Authenticate header per {{RFC9110}}. The response body SHOULD include a generic error indicator without revealing which specific verification step failed. The receiving agent MUST NOT process the requested action when EAT verification fails.

DAG Validation

EATs form a Directed Acyclic Graph (DAG) where each task references its parent tasks via the "pred" claim. DAG validation is performed against the EAT store — either an audit ledger or the set of parent EATs received inline.

DAG validation applies regardless of assurance level.

When receiving and verifying an EAT, implementations MUST perform the following DAG validation steps:

1. Task ID Uniqueness: The "jti" claim MUST be unique within the applicable scope (the workflow identified by "wid", or the entire EAT store if "wid" is absent). If an EAT with the same "jti" already exists, the EAT MUST be rejected.

2. Parent Existence: Every task identifier listed in the "pred" array MUST correspond to a task that is available in the EAT store (either previously recorded in the ledger or received inline as a verified parent EAT). If any parent task is not found, the EAT MUST be rejected.

3. Temporal Ordering: The "iat" value of every parent task MUST NOT be greater than the "iat" value of the current task plus a configurable clock skew tolerance (RECOMMENDED: 30 seconds). That is, for each parent: parent.iat < child.iat + clock_skew_tolerance. The tolerance accounts for clock skew between agents; it does not guarantee strict causal ordering from timestamps alone. Causal ordering is primarily enforced by the DAG structure (parent existence in the EAT store), not by timestamps. If any parent task violates this constraint, the EAT MUST be rejected.

4. Acyclicity: Following the chain of parent references MUST NOT lead back to the current EAT's "jti". If a cycle is detected, the EAT MUST be rejected. Note: because the Parent Existence check (step 2) requires that all parents already exist in the EAT store, and an EAT cannot reference itself or a future EAT, cycles are prevented by construction. This explicit check serves as defense in depth against implementation errors or store corruption.

5. Domain Consistency: When "domain" is present, parent tasks SHOULD belong to the same domain or to a domain with which a federation relationship has been established. Cross-domain parent references are valid when the domains have a pre-established trust relationship.

6. Workflow Consistency: When "wid" is present, verifiers SHOULD validate that the "wid" of each parent EAT matches the "wid" of the current EAT. Cross-workflow parent references (where a parent's "wid" differs from the child's "wid") MUST be rejected unless explicitly permitted by deployment policy.

7. Cross-Workflow Composition: When a task in one workflow depends on the output of a task in a different workflow, the dependency MAY be expressed as a cross-workflow reference using a composite identifier in the "pred" array: <wid>:<jti>. Cross-workflow references MUST be explicitly enabled by deployment policy and are subject to all verification requirements of both the source and target workflows.

To prevent denial-of-service via extremely deep or wide DAGs, implementations SHOULD enforce a maximum ancestor traversal limit (RECOMMENDED: 10000 nodes). If the limit is reached before cycle detection completes, the EAT SHOULD be rejected.

In distributed deployments, a parent EAT may not yet be available locally due to replication lag. Implementations MAY defer validation to allow parent EATs to arrive, but MUST NOT treat the EAT as verified until all parent references are resolved.

Audit Ledger Interface

EATs MAY be recorded in an immutable audit ledger for compliance verification and post-hoc analysis. A ledger is OPTIONAL for L1 and L2 deployments but is REQUIRED for L3 deployments. This specification does not mandate a specific storage technology. Implementations MAY use append-only logs, databases with cryptographic commitment schemes, distributed ledgers, or any storage mechanism that provides the required properties.

Baseline Ledger Requirements

When an audit ledger is deployed, the implementation MUST provide:

1. Append-only semantics: Once an EAT is recorded, it MUST NOT be modified or deleted.

2. Ordering: The ledger MUST maintain a total ordering of EAT entries via a monotonically increasing sequence number.

3. Lookup by EAT ID: The ledger MUST support efficient retrieval of EAT entries by "jti" value.

4. Lookup by Workflow ID: The ledger SHOULD support retrieval of all EAT entries matching a given "wid" value.

5. Integrity verification: The ledger SHOULD provide a mechanism to verify that no entries have been tampered with (e.g., hash chains or Merkle trees).

The ledger SHOULD be maintained by an entity independent of the workflow agents to reduce the risk of collusion.

L3 Ledger Requirements

At Level 3, the audit ledger MUST satisfy the baseline requirements above and the additional requirements defined in {{l3-ledger}}. In particular:

• Each entry MUST include a cryptographic hash of the previous entry, forming a verifiable hash chain.

• The ledger MUST provide cryptographic commitment proofs (Merkle tree proofs, signed hash chain roots, or equivalent).

• The ledger MUST return a receipt upon successful append, containing the sequence number, entry hash, and commitment proof.

Ledger Independence

The audit ledger SHOULD be operated by an entity that is independent of the workflow agents. When the ledger operator and the workflow agents are controlled by the same organization, additional safeguards SHOULD be implemented, such as separation of duties between ledger administrators and agent operators, or independent ledger replicas maintained by third parties.

EAT Discovery API

Deployments SHOULD provide a standardized API for EAT discovery and verification. The following endpoints are RECOMMENDED:

Lookup by Task ID:

GET /eat/{jti} — Returns the EAT and its ledger receipt (if recorded) for the specified task identifier.

Lookup by Workflow ID:

GET /eat?wid={wid} — Returns all EATs in the specified workflow, ordered by ledger sequence number.

DAG Retrieval:

GET /eat/{jti}/dag — Returns the complete DAG rooted at the specified task, including all ancestor EATs.

Ledger Verification:

GET /ledger/verify?seq={seq} — Returns the commitment proof for the specified ledger sequence number.

These endpoints MUST require authentication and MUST enforce access control per the deployment's authorization policy.

L3 Ledger Verification

Auditors verifying L3 ledger integrity SHOULD perform the following checks:

1. Verify the hash chain: for each entry, recompute the hash of the previous entry and confirm it matches the recorded value.

2. Verify the cryptographic commitment: validate Merkle inclusion proofs (or equivalent) for a sample or all entries.

3. Verify EAT signatures: for each EAT in the ledger, perform the L2 verification procedure ({{l2-verification}}) using the public keys valid at the time of recording.

4. Verify DAG consistency: confirm that all parent references in all EATs resolve to valid entries in the ledger.

Security Considerations

Threat Model

The threat model considers: (1) a malicious agent that creates false EAT claims, (2) an agent whose private key has been compromised, (3) a ledger tamperer attempting to modify recorded entries, (4) a time manipulator altering timestamps to affect perceived ordering, (5) a domain boundary attacker attempting to inject, modify, or drop EATs at domain transitions, and (6) a regulatory arbitrage attacker claiming false regulatory profiles to avoid compliance requirements.

Level-Specific Security Properties

Level 1

L1 provides no cryptographic binding between the EAT and its issuer. A compromised or malicious intermediary with access to the transport channel can modify L1 EATs without detection after delivery. L1 does not provide non-repudiation: the issuer can deny having produced an EAT, and the receiver cannot prove otherwise.

L1 MUST NOT be used across trust domain boundaries. L1 is appropriate only when all agents are operated by the same organization, the transport channel is fully trusted (e.g., service mesh with mTLS), and the deployment does not require non-repudiation or tamper evidence beyond transport security.

Level 2

L2 inherits all security properties of the JWS-based EAT mechanism defined in this document. The existing security analysis (signature verification, replay prevention, key compromise, collusion) applies directly to L2.

Level 3

L3 provides all L2 security properties plus tamper-evident history via the audit ledger's hash chain and cryptographic commitment. Modifications to ledger entries are detectable through hash chain verification. Deletions or insertions are detectable through Merkle proof validation.

L3 does not prevent a compromised ledger operator from refusing to record entries (censorship), but the submitting agent's receipt mechanism ({{l3-ledger}}) provides evidence of submission. Deployments concerned about ledger censorship SHOULD use multiple independent ledger replicas.

Assurance Level Downgrade Attacks

The assurance level of an EAT is determined by its format: unsigned JSON indicates L1, while a JWS compact serialization indicates L2 or L3. This format-based detection determines what was sent, not what was expected by the verifier.

A man-in-the-middle or compromised proxy could strip the JWS signature from an L2 or L3 EAT and re-encode the payload as an unsigned L1 JSON object. Because the resulting EAT is syntactically valid at L1, a verifier that accepts any assurance level would process it without detecting the downgrade.

Verifiers MUST be configured with a minimum acceptable assurance level and MUST reject EATs whose detected level falls below that minimum. Format-based level detection alone is insufficient without a policy-enforced minimum level requirement.

Cross-Domain Security Considerations

Domain Boundary Attacks

At domain boundaries, additional attack vectors exist:

• EAT injection: An attacker at the boundary inserts a fabricated EAT into the workflow. Mitigation: L2/L3 signature verification prevents injection of unsigned or incorrectly signed EATs.

• EAT dropping: An attacker removes an EAT at the boundary, breaking the DAG chain. Mitigation: the receiving agent's DAG validation (parent existence check) detects missing parents.

• Regulatory profile falsification: An agent claims a regulatory profile that does not actually apply, to avoid stricter compliance requirements. Mitigation: receiving domains SHOULD independently verify that the claimed jurisdiction and regulatory profiles are consistent with the sending domain's known regulatory environment.

Mixed Identity Framework Security

When a DAG spans agents using different identity frameworks (e.g., X.509 in one domain, OAuth in another), the verifier must support multiple key resolution mechanisms. This increases the attack surface:

• Each identity framework has its own revocation semantics and timing. A key revoked in one framework may not be immediately detectable through another framework's mechanisms.

• Deployments SHOULD maintain a unified revocation status cache that aggregates revocation information across all supported identity frameworks.

Federated Ledger Security

• Split-view attacks: A foreign ledger could present different receipts to different verifiers. Mitigation: cross-ledger anchoring includes the commitment proof, which can be independently verified.

• Receipt replay: A valid receipt from one EAT could be attached to a different EAT. Mitigation: the "eat_hash" field in the receipt binds the receipt to a specific EAT.

• Ledger key compromise: If a foreign ledger's signing key is compromised, fake receipts can be generated. Mitigation: ledger keys SHOULD be stored in HSMs; receipts SHOULD be cross-checked against the foreign ledger's published commitment history.

Self-Assertion Limitation

EATs are self-asserted by the executing agent. The agent claims what it did, and this claim is signed with its private key (at L2 and L3). A compromised or malicious agent could create EATs with false claims (e.g., claiming an action was performed when it was not).

EATs do not independently verify that:

• The claimed execution actually occurred as described
• The input/output hashes correspond to the actual data processed
• The agent faithfully performed the stated action
• The claimed regulatory profile is accurate

The trustworthiness of EAT claims depends on the trustworthiness of the signing agent and the integrity of the broader deployment environment. EATs provide a technical mechanism for execution recording; they do not by themselves satisfy any specific regulatory compliance requirement.

Signature Verification

For L2 and L3 deployments, EATs MUST be signed with the agent's private key using JWS {{RFC7515}}. The signature algorithm MUST match the algorithm associated with the agent's identity credential (see {{identity-binding}}). Receivers MUST verify the EAT signature against the corresponding public key before processing any claims. Receivers MUST verify that the signing key has not been revoked within the trust domain (see step 6 in {{l2-verification}}).

If signature verification fails or if the signing key has been revoked, the EAT MUST be rejected entirely and the failure MUST be logged.

Implementations MUST use established JWS libraries and MUST NOT implement custom signature verification. Implementations SHOULD follow the JWT security best practices defined in {{RFC8725}}.

Replay Attack Prevention

EATs include short expiration times (RECOMMENDED: 5-15 minutes) and audience restriction via "aud" to limit replay attacks. Implementations MUST maintain a cache of recently-seen "jti" values and MUST reject EATs with duplicate "jti" values. For L2 and L3, each EAT is cryptographically bound to the issuing agent via "kid"; verifiers MUST confirm that "kid" resolves to the "iss" agent's key (step 8 in {{l2-verification}}).

Man-in-the-Middle Protection

EATs MUST be transmitted over TLS or mTLS connections.

Key Compromise

If an agent's private key is compromised, an attacker can forge EATs that appear to originate from that agent. Mitigations:

• Implementations SHOULD use short-lived keys and rotate them frequently.
• Private keys SHOULD be stored in hardware security modules or equivalent secure key storage.
• Trust domains MUST support rapid key revocation.

EATs recorded before key revocation remain valid historical records but SHOULD be flagged for audit purposes. New EATs MUST NOT reference a parent EAT whose signing key is known to be revoked at creation time.

Collusion and DAG Integrity

A single malicious agent cannot forge parent task references because DAG validation requires parent tasks to exist in the EAT store. However, multiple colluding agents could create a false execution history. Additionally, a malicious agent may omit actual parent dependencies from "pred" to hide influences on its output; because EATs are self-asserted ({{self-assertion-limitation}}), no mechanism can force complete dependency declaration.

Mitigations include:

• The ledger SHOULD be maintained by an entity independent of the workflow agents.
• Multiple independent ledger replicas can be compared for consistency.
• External auditors can compare the declared DAG against expected workflow patterns.

Privilege Escalation via EATs

EATs record execution history; they do not convey authorization. Verifiers MUST NOT interpret the presence of an EAT, or a particular set of parent references in "pred", as an authorization grant. Authorization decisions MUST remain with the deployment's identity and authorization layer.

Denial of Service

Implementations SHOULD apply rate limiting to prevent excessive EAT submissions. For L2 and L3, DAG validation SHOULD be performed after signature verification to avoid wasting resources on unsigned or incorrectly signed tokens. L1 EATs are cheaper to validate (no signature verification) but are correspondingly easier for an attacker to generate; L1 deployments SHOULD apply stricter rate limits to compensate.

Timestamp Accuracy

Implementations SHOULD use synchronized time sources (e.g., NTP) and SHOULD allow a configurable clock skew tolerance (RECOMMENDED: 30 seconds). Cross-domain deployments MAY require a higher tolerance and SHOULD document the configured value.

EAT Size Constraints

Implementations SHOULD limit the "pred" array to a maximum of 256 entries. See {{extension-claims}} for "eat_ext" size limits.

When EATs are transported via HTTP headers, the total encoded size of the Execution-Context header value is subject to practical limits imposed by HTTP servers and intermediaries. Many implementations enforce header size limits of 8 KB or 16 KB. Implementations SHOULD ensure that the total size of an EAT (including JWS overhead for L2/L3) does not exceed 8 KB when transported via HTTP header. EATs that exceed HTTP header size limits SHOULD be transported in the HTTP request body instead (see {{l1-transport}} and {{l2-transport}}).

Identity Binding Security

JWK Set Binding

When identity is bound via JWK Set URI (see {{identity-binding}}), there is a time-of-check/time-of-use gap between JWK set refreshes. A key that has been removed from the JWK set may still be accepted by a verifier whose cached copy has not yet been refreshed. Implementations SHOULD refresh JWK sets at configurable intervals (RECOMMENDED: no longer than 5 minutes).

X.509 Binding

When identity is bound via X.509 certificates, revocation checking depends on OCSP responder or CRL distribution point availability. If the revocation source is unreachable, the verifier needs to decide whether to accept or reject the EAT. Implementations SHOULD hard-fail for L3 (reject the EAT if revocation status cannot be determined), as L3 workflows require the strongest assurance. Implementations MAY soft-fail for L2 with logging, accepting the EAT but recording the revocation check failure for subsequent audit review.

Audit Ledger Threats

Availability

If the audit ledger is unavailable in synchronous recording mode (see {{l3-ledger}}), all L3 workflows halt because agents cannot obtain ledger receipts. Deployments SHOULD implement ledger redundancy (e.g., multiple ledger replicas behind a load balancer) to prevent the ledger from becoming a single point of failure.

Split-View Attacks

A compromised ledger could present different views to different verifiers (equivocation), causing inconsistent audit state across the deployment. Deployments SHOULD use multiple independent ledger replicas and SHOULD periodically compare their state to detect divergence.

Receipt Authenticity

If the ledger's signing key is compromised, an attacker can generate fake receipts for entries that were never recorded. Ledger signing keys SHOULD be stored in hardware security modules (HSMs) and SHOULD be rotated regularly.

Asynchronous Recording Gap

In asynchronous recording mode (see {{l3-ledger}}), downstream agents act on EATs before ledger confirmation is received. During this gap, an EAT that will ultimately fail ledger recording may already have influenced downstream workflow steps. Deployments using asynchronous recording SHOULD implement reconciliation procedures to detect and handle EATs that fail ledger confirmation after downstream processing has begun.

Privacy Considerations

Data Exposure in EATs

EATs necessarily reveal:

• Agent identities ("iss", "aud") for accountability purposes
• Action descriptions ("exec_act") for audit trail completeness
• Timestamps ("iat", "exp") for temporal ordering
• Domain information ("domain") for regulatory routing

EATs are designed to NOT reveal:

• Actual input or output data values (replaced with cryptographic hashes via "inp_hash" and "out_hash")
• Internal computation details or intermediate steps
• Proprietary algorithms or intellectual property
• Personally identifiable information (PII)

Privacy exposure is equivalent across all assurance levels: the EAT payload contains the same claims regardless of whether the EAT is unsigned (L1), signed (L2), or signed with ledger recording (L3). L3 ledger recording increases the durability of the privacy-relevant data but does not increase its scope.

Cross-Domain Privacy

Cross-domain workflows introduce additional privacy concerns:

• Workflow topology leakage: The DAG structure reveals which agents interact across domains, potentially exposing organizational relationships or business processes.

• Regulatory conflicts: GDPR's data minimization requirements may conflict with another jurisdiction's audit trail completeness requirements. The selective disclosure mechanism ({{selective-disclosure}}) helps resolve this tension.

• Jurisdiction hopping: EATs that cross multiple jurisdictions accumulate metadata. Deployments SHOULD use selective disclosure to minimize cross-jurisdictional data exposure.

Data Minimization

Implementations SHOULD minimize the information included in EATs. The "exec_act" claim SHOULD use structured identifiers (e.g., "process_payment") rather than natural language descriptions. Extension keys in "eat_ext" ({{extension-claims}}) deserve particular attention: human-readable values risk exposing sensitive operational details.

Storage and Access Control

EATs stored in audit ledgers SHOULD be access-controlled so that only authorized auditors can read them. Implementations SHOULD consider encryption at rest for ledger storage. EATs provide structural records of execution ordering; they are not intended for public disclosure.

Full input and output data (corresponding to the hashes in EATs) SHOULD be stored separately from the ledger with additional access controls, since auditors may need to verify hash correctness but general access to the data values is not needed.

Workflow Topology Leakage

The DAG structure of EATs reveals workflow topology: which agents interact, fan-out and fan-in patterns, sequential versus parallel execution, and organizational structure. At L3, this topology is permanently recorded in the audit ledger. Deployments SHOULD consider whether workflow topology constitutes sensitive information and apply appropriate access controls to EAT stores and ledgers.

Cross-Workflow Correlation

Stable agent identifiers in the "iss" claim enable cross-workflow activity correlation: an observer with access to EATs from multiple workflows can track which agents participate in which workflows and how frequently. Deployments with privacy requirements MAY use per-workflow or rotating agent identifiers where feasible to limit cross-workflow correlation.

IANA Considerations

Media Type Registration

This document requests registration of the following media type in the "Media Types" registry maintained by IANA:

Note: The media type "application/eat+jwt" uses the "+jwt" structured syntax suffix. While "+jwt" is widely used in practice, it is not yet a formally registered structured syntax suffix per {{RFC6838}}. Registration of the "+jwt" suffix is the subject of ongoing work in the IETF.

application/eat+jwt

Type name:

application

Subtype name:

eat+jwt

Required parameters:

none

Optional parameters:

none

Encoding considerations:

8bit; at Level 2 and Level 3, an EAT is a JWT that is a JWS using the Compact Serialization, which is a sequence of Base64url-encoded values separated by period characters. At Level 1, this media type is not used; L1 EATs use application/json.

Security considerations:

See the Security Considerations section of this document.

Interoperability considerations:

none

Published specification:

This document

Applications that use this media type:

Applications that implement agentic workflows requiring cross-domain execution audit trails.

Additional information:

Magic number(s): none File extension(s): none Macintosh file type code(s): none

Person and email address to contact for further information:

Christian Nennemann, ietf@nennemann.de

Intended usage:

COMMON

Restrictions on usage:

none

Author:

Christian Nennemann

Change controller:

IETF

HTTP Header Field Registration

This document requests registration of the following header field in the "Hypertext Transfer Protocol (HTTP) Field Name Registry" maintained by IANA:

Field name:

Execution-Context

Status:

permanent

Specification document:

This document, {{http-header}}

JWT Claims Registration

This document requests registration of the following claims in the "JSON Web Token Claims" registry maintained by IANA:

Claim Name	Claim Description	Change Controller	Reference
wid	Workflow Identifier	IETF	{{exec-claims}}
exec_act	Action/Task Type	IETF	{{exec-claims}}
pred	Predecessor Task Identifiers	IETF	{{exec-claims}}
actor_type	Actor Type	IETF	{{actor-type-claim}}
inp_hash	Input Data Hash	IETF	{{data-integrity-claims}}
out_hash	Output Data Hash	IETF	{{data-integrity-claims}}
domain	Trust Domain Identifier	IETF	{{domain-claims}}
eat_reg	Regulatory Profile Metadata	IETF	{{domain-claims}}
ledger_receipt	Audit Ledger Receipt	IETF	{{domain-claims}}
eat_ext	Extension Object	IETF	{{extension-claims}}
{: #table-claims title="JWT Claims Registrations"}

Regulatory Profile Registry

This document establishes a new "EAT Regulatory Profiles" registry under the "JSON Web Token (JWT)" group. Registration policy is Specification Required per RFC 8126.

Each registration includes:

• Profile URI (e.g., "urn:ietf:eat:reg:eu-dora")
• Description
• Required EAT claims
• Minimum assurance level
• Reference document

Initial registrations are defined in {{profile-registry}}.

--- back

Use Cases

{:numbered="false"}

This section describes representative use cases demonstrating how EATs provide structured execution records across multiple domains with different regulatory requirements.

Note: task identifiers in this section are abbreviated for readability. In production, all "jti" values are required to be UUIDs per {{exec-claims}}.

Cross-Jurisdiction Financial Trading (L3)

{:numbered="false"}

An investment bank in Frankfurt coordinates with a credit rating agency in New York. The bank operates under DORA and MiFID II; the rating agency operates under US SOC 2. The DAG spans both jurisdictions with different regulatory profiles.

Domain: bank.example.com (EU, DORA + MiFID II)
  Agent A1 (Portfolio Risk):
    jti: task-001    pred:[]
    iss: https://agents.bank.example.com/risk
    exec_act: analyze_portfolio_risk
    domain: bank.example.com
    eat_reg: {profiles: ["urn:ietf:eat:reg:eu-dora",
              "urn:ietf:eat:reg:financial-mifid2"],
              retention_days: 2555, jurisdiction: "EU"}

Domain: ratings.example.com (US, SOC 2)
  Agent B1 (Credit Rating):
    jti: task-002    pred:[]
    iss: https://agents.ratings.example.com/credit
    exec_act: assess_credit_rating
    domain: ratings.example.com
    eat_reg: {profiles: ["urn:ietf:eat:reg:us-soc2"],
              retention_days: 365, jurisdiction: "US"}

Domain: bank.example.com (EU)
  Agent A2 (Compliance):
    jti: task-003    pred:[task-001, task-002]
    iss: https://agents.bank.example.com/compliance
    exec_act: verify_trade_compliance
    domain: bank.example.com
    eat_reg: {profiles: ["urn:ietf:eat:reg:eu-dora"],
              retention_days: 2555, jurisdiction: "EU"}

  Agent A3 (Execution):
    jti: task-004    pred:[task-003]
    iss: https://agents.bank.example.com/execution
    exec_act: execute_trade

{: #fig-finance title="Cross-Jurisdiction Trading (L3)"}

The resulting DAG:

task-001 (analyze_portfolio_risk)   task-002 (assess_credit_rating)
  [bank.example.com, EU]             [ratings.example.com, US]
          \                              /
           v                            v
         task-003 (verify_trade_compliance)
           [bank.example.com, EU]
                    |
                    v
         task-004 (execute_trade)
           [bank.example.com, EU]

{: #fig-finance-dag title="Cross-Jurisdiction DAG"}

Task 002 from the US domain carries a ledger_receipt from the US ledger. When task 003 in the EU domain references task 002 as a parent, the EU ledger records the cross-ledger anchor, creating an end-to-end verifiable audit trail across both jurisdictions.

Healthcare Agent Pipeline with HITL (L2/L3)

{:numbered="false"}

A clinical AI agent recommends a treatment plan that requires human physician approval before the pharmacy agent can proceed. The workflow demonstrates the "actor_type" claim for human-in-the-loop steps.

Domain: hospital.example.org (US, HIPAA)

  Agent C1 (Clinical AI):
    jti: task-101    pred:[]
    iss: https://agents.hospital.example.org/clinical-ai
    exec_act: recommend_treatment
    actor_type: agent
    eat_reg: {profiles: ["urn:ietf:eat:reg:us-hipaa"],
              retention_days: 2190, jurisdiction: "US"}

  Dr. Smith (Physician Approval):
    jti: task-102    pred:[task-101]
    iss: https://staff.hospital.example.org/dr-smith
    exec_act: approve_treatment
    actor_type: human
    eat_reg: {profiles: ["urn:ietf:eat:reg:us-hipaa"],
              retention_days: 2190, jurisdiction: "US"}

  Agent C2 (Pharmacy):
    jti: task-103    pred:[task-102]
    iss: https://agents.hospital.example.org/pharmacy
    exec_act: dispense_medication
    actor_type: agent

{: #fig-healthcare title="Healthcare HITL Workflow (L2/L3)"}

Task 102 has "actor_type": "human", indicating a physician reviewed and approved the AI's recommendation. The EAT is signed by the hospital's approval gateway on behalf of Dr. Smith. When this workflow's output crosses to a research partner under GDPR, selective disclosure redacts the physician's identity and treatment details while preserving the DAG structure proving human approval occurred.

Multi-Cloud AI Pipeline with Mixed Identity (L2)

{:numbered="false"}

An AI pipeline spans three cloud providers, each using a different identity framework. The DAG demonstrates cross-domain identity resolution.

Domain: aws.customer.example (X.509 identity)
  Agent P1 (Data Preprocessor):
    jti: task-301    pred:[]
    iss: urn:x509:CN=preprocessor.aws.customer.example
    exec_act: preprocess_training_data
    kid: x5t#sha256:abc123...

Domain: gcp.partner.example (OAuth identity)
  Agent P2 (Model Training):
    jti: task-302    pred:[task-301]
    iss: https://oauth.gcp.partner.example/client/trainer
    exec_act: train_model
    kid: oauth-key-456

Domain: azure.customer.example (JWK Set identity)
  Agent P3 (Model Validation):
    jti: task-303    pred:[task-302]
    iss: https://agents.azure.customer.example/validator
    exec_act: validate_model
    kid: jwk-789

{: #fig-multicloud title="Multi-Cloud Mixed Identity (L2)"}

Each domain uses its own identity framework; the EAT format remains the same across all three. Cross-domain key resolution is handled through pre-established federation (e.g., shared JWK set endpoints).

Internal Microservice Mesh (L1)

{:numbered="false"}

In an internal AI platform, multiple microservices coordinate to process requests. All agents operate within a single trust domain behind a service mesh with mTLS. Non-repudiation is not required; the EATs are used for observability and debugging.

Domain: internal.example
  Agent S1 (Preprocessor):
    jti: task-401    pred:[]
    exec_act: preprocess_input

  Agent S2 (Model Inference):
    jti: task-402    pred:[task-401]
    exec_act: run_inference

  Agent S3 (Postprocessor):
    jti: task-403    pred:[task-402]
    exec_act: format_output

{: #fig-internal title="Internal Microservice Workflow (L1)"}

Relationship to WIMSE ECT

{:numbered="false"}

This specification generalizes the Execution Context Token (ECT) concept originally defined in the context of the WIMSE working group. The WIMSE ECT specification focuses on SPIFFE/WIMSE-based identity; this specification provides the same core functionality (DAG-based execution audit with three assurance levels) while being identity-framework agnostic.

Deployments using WIMSE can implement EATs with the following profile:

• "iss" is the workload's SPIFFE ID (e.g., spiffe://trust-domain/path)
• "kid" references the WIT public key
• "alg" matches the WIT algorithm
• Revocation is checked via SPIFFE trust bundles
• EATs are transported alongside WIT and WPT per the WIMSE service-to-service protocol {{I-D.ietf-wimse-s2s-protocol}}

Related Work

{:numbered="false"}

OAuth 2.0 Token Exchange and the "act" Claim

{:numbered="false"}

{{RFC8693}} defines the OAuth 2.0 Token Exchange protocol and registers the "act" (Actor) claim in the JWT Claims registry. The "act" claim creates nested JSON objects representing a delegation chain: "who is acting on behalf of whom." While the nesting superficially resembles a chain, it is strictly linear (each "act" object contains at most one nested "act"), represents authorization delegation rather than task execution, and carries no task identifiers or input/output integrity data. The "act" chain cannot represent branching (fan-out) or convergence (fan-in) and therefore cannot form a DAG.

EATs intentionally use the distinct claim name "exec_act" for the action/task type to avoid collision with the "act" claim.

Transaction Tokens

{:numbered="false"}

OAuth Transaction Tokens {{I-D.ietf-oauth-transaction-tokens}} propagate authorization context across workload call chains. The Txn-Token "req_wl" claim accumulates a comma-separated list of workloads that requested replacement tokens, which is the closest existing mechanism to call-chain recording.

However, "req_wl" cannot form a DAG because it is linear, it is incomplete (only workloads that request replacement tokens are recorded), and it cannot represent convergence (fan-in).

EATs and Transaction Tokens are complementary: a Txn-Token propagates authorization context ("this request is authorized for scope X on behalf of user Y"), while an EAT records execution accountability ("task T was performed, depending on tasks P1 and P2").

Distributed Tracing (OpenTelemetry)

{:numbered="false"}

OpenTelemetry {{OPENTELEMETRY}} and similar distributed tracing systems provide observability for debugging and monitoring. EATs differ in several important ways: at L2 and L3, EATs are cryptographically signed per-task with the agent's private key; EATs are tamper-evident through JWS signatures; EATs enforce DAG validation rules; and EATs are designed for regulatory audit rather than operational monitoring. OpenTelemetry data is typically controlled by the platform operator and can be modified or deleted without detection.

SCITT (Supply Chain Integrity, Transparency, and Trust)

{:numbered="false"}

The SCITT architecture {{I-D.ietf-scitt-architecture}} defines a framework for transparent and auditable supply chain records. There is a notable parallel between SCITT's Transparency Service and EAT's Level 3 audit ledger: both use append-only logs with cryptographic commitment. The federated ledger model defined in this specification ({{federated-ledger}}) draws on similar principles to Certificate Transparency {{CERTIFICATE-TRANSPARENCY}} and SCITT.

RATS (Remote Attestation Procedures)

{:numbered="false"}

RATS {{RFC9334}} defines an architecture for conveying attestation evidence about platform trustworthiness. RATS attests "is this platform in a trustworthy state?" while EATs record "what did this agent do?" — both deal with claims about entities but at different layers. EATs could complement RATS by recording execution context on platforms whose trustworthiness has been established through RATS attestation.

Emerging Agent Protocol Frameworks

{:numbered="false"}

Several emerging frameworks address agent-to-agent communication, including Google's Agent-to-Agent Protocol (A2A), Anthropic's Model Context Protocol (MCP), and orchestration frameworks such as LangChain and LangGraph. These frameworks primarily address agent discovery, message routing, and tool invocation but do not provide cryptographically verifiable execution records or DAG-based audit trails. EATs complement these frameworks by adding an execution accountability layer.

Acknowledgments

{:numbered="false"}

The author thanks the WIMSE working group for their foundational work on workload identity in multi-system environments, which provided the original motivation for execution context tracing.