<?xml version="1.0" encoding="utf-8"?>
<!DOCTYPE rfc [
  <!ENTITY nbsp "&#160;">
  <!ENTITY zwsp "&#8203;">
  <!ENTITY nbhy "&#8209;">
  <!ENTITY wj "&#8288;">
]>
<rfc xmlns:xi="http://www.w3.org/2001/XInclude"
     category="info"
     docName="draft-das-ai-agent-cyberattack-effect-finality-00"
     ipr="trust200902"
     submissionType="independent"
     xml:lang="en"
     tocInclude="true"
     tocDepth="3"
     symRefs="true"
     sortRefs="true"
     version="3">

<front>
  <title abbrev="AI Agent Effect Finality">AI Safety: When AI Agents Initiate Cyberattacks -- Execution Finality for Data Exfiltration, Payment Diversion, Infrastructure Changes, and Code-Executing Agents</title>
  <seriesInfo name="Internet-Draft" value="draft-das-ai-agent-cyberattack-effect-finality-00"/>

  <author fullname="Sangam Das" initials="S." surname="Das">
    <organization>Independent</organization>
    <address>
      <postal>
        <city>Balasore</city>
        <region>Odisha</region>
        <country>India</country>
      </postal>
      <phone>+91-9861363532</phone>
      <email>info@sangamdas.com</email>
    </address>
  </author>

  <date year="2026" month="November" day="17"/>
  <area>Security</area>
  <workgroup>Independent Submission</workgroup>
  <keyword>AI safety</keyword>
  <keyword>AI agents</keyword>
  <keyword>execution finality</keyword>
  <keyword>prompt injection</keyword>
  <keyword>policy enforcement point</keyword>
  <keyword>finality sink</keyword>
  <keyword>agent security</keyword>

  <abstract>
    <t>Problem. An AI agent can cause an unauthorized real-world effect -- a
    confidential data export, a payment to a substituted beneficiary, a
    network or industrial configuration change, a code deployment or a
    secret release -- while every individual check succeeds: the model
    accepted a plausible task, the agent held a valid credential, the tool
    call was syntactically permitted, and the gateway saw an allowed
    method. The gap is that the check and the effect are not the same
    thing. Identity and access controls answer "may this principal call
    this interface?"; they do not, by themselves, answer "may this exact
    effect, with these resolved targets and values, complete now, on this
    route, given what has already happened?"</t>
    <t>Solution. This document describes an execution-finality profile for
    AI-initiated operations. An agent's proposal is held as a non-effective
    Candidate Act. A protected validator compares a canonical descriptor of
    the operation actually about to be committed with a source-authority
    envelope that model-editable content cannot create or enlarge. A
    Finality Sink, which owns the effect-causing credential or commit right,
    verifies an act-bound, non-bearer Execution Handle over the same bytes
    it commits, consumes it atomically with the commit, and emits an
    authenticated destination receipt. Later phases and retries are gated on
    that receipt rather than on the agent's own report. Unknown outcomes
    become INDETERMINATE and are reconciled before any retry.</t>
    <t>Context. The document analyses four scenarios (enterprise data
    exfiltration, payment diversion, critical-infrastructure configuration
    manipulation, and code-executing agents), specifies data objects, a
    Candidate Act state machine, atomic consumption and crash-recovery
    rules, sink trust, delegation attenuation, protocol placement for tool
    and agent-to-agent calls, revocation freshness, latency placement,
    legacy integration and a falsifiable evaluation plan. It is a
    prospective architecture: no measured latency, blocking rate or
    deployment is claimed. A reviewer FAQ is included. Critical technical
    review is invited.</t>
  </abstract>
</front>

<middle>

<section anchor="intro" numbered="true">
  <name>Introduction</name>

  <section numbered="true">
    <name>The Problem Space</name>
    <t>Documented agent risks include indirect prompt injection through
    retrieved content, tool misuse, excessive agency, and AI-orchestrated
    intrusion activity in which humans select targets while agents perform
    most operational steps <xref target="OWASP-AGENTIC"/>
    <xref target="NIST-AI-100-2"/> <xref target="ANTHROPIC-ESPIONAGE"/>
    <xref target="MITRE-T1698"/>. The consequential event in each case is
    not model text. It is a committed effect: bytes leaving an
    organization, funds moving, a rule installed in a controller, a
    commit merged, a secret released.</t>
    <t>Mature deployments already constrain many of these effects with
    identity and access management, Zero Trust policy enforcement points
    <xref target="NIST-SP-800-207"/>, attribute-based access control
    <xref target="NIST-SP-800-162"/>, data-loss prevention, payment risk
    engines, change management and operational-technology safety systems
    <xref target="NIST-SP-800-82"/>. This document does not claim that
    such controls are inherently bypassable. It identifies four narrower
    residual gaps that appear in partial deployments:</t>
    <ol>
      <li>After authorization, before effect: the target, amount, scope or
      destination changes between an approval and the operation actually
      committed.</li>
      <li>After a first effect, before continuation: reported success is
      false, late or missing, and the agent continues or retries on its own
      account.</li>
      <li>Across routes: the same outcome is reachable through a direct API,
      a queue consumer, an administrative interface, a hosted tool or a
      delegated agent that does not pass the expected gate.</li>
      <li>Across transformations: a benign-looking entry operation causes
      material transitive effects through stored procedures, workflow rules,
      controller reconciliation or downstream fan-out.</li>
    </ol>
  </section>

  <section numbered="true">
    <name>Why Common Protocol Work Is Needed</name>
    <t>Each of these gaps can be closed locally by a single vendor for a
    single resource. The cross-vendor problem is that an agent's operation
    typically crosses an orchestrator, a tool server, an identity provider,
    a gateway and a destination owned by different parties. Without a
    shared representation of "the act that was authorized", a shared rule
    for binding authority to that act at the commit point, and a shared form
    for destination evidence, each hop re-decides on partial information.
    The objects and rules in this document are intended as a basis for that
    shared representation; they reuse existing IETF mechanisms where they
    fit (<xref target="ietf-relation"/>).</t>
  </section>

  <section numbered="true">
    <name>What Is New in This Document</name>
    <t>Relative to the companion documents listed in
    <xref target="companions"/>, this document contributes:</t>
    <ul>
      <li>an adversary model and four end-to-end scenarios specific to
      AI-initiated attacks, each with a best-practice conventional baseline
      for comparison;</li>
      <li>an explicit rule that the sink verifies a digest computed over the
      final bytes it is about to commit (time-of-check to time-of-use
      closure, <xref target="toctou"/>);</li>
      <li>a Candidate Act state machine with permitted transitions and the
      component allowed to trigger each (<xref target="state-machine"/>);</li>
      <li>crash-consistency rules for atomic consumption across shards and
      failover (<xref target="consume"/>);</li>
      <li>delegation attenuation rules for multi-agent systems
      (<xref target="delegation"/>);</li>
      <li>placement for tool calls and agent-to-agent calls, and a
      comparison with OAuth Rich Authorization Requests, Token Exchange,
      DPoP and Transaction Tokens (<xref target="protocol-placement"/>);</li>
      <li>a bounded-staleness model for revocation and offline leases
      (<xref target="revocation"/>);</li>
      <li>a code-executing-agent scenario and an output-channel (response
      egress) profile (<xref target="scenario-d"/>,
      <xref target="output-channel"/>);</li>
      <li>a falsifiable evaluation plan against a strong conventional
      baseline (<xref target="evaluation"/>) and a reviewer FAQ
      (<xref target="faq"/>).</li>
    </ul>
  </section>

  <section numbered="true">
    <name>Scope and Non-Goals</name>
    <t>In scope: consequential operations that have a narrow protected
    commit point -- data export or publication, payment initiation, capture
    and settlement phases, privileged configuration activation, credential
    or secret release, code merge and deployment, and command activation.</t>
    <t>Not in scope: deciding whether a model should follow an instruction;
    validating every token, packet, database row or radio/actuator control
    cycle; replacing safety-certified interlocks; controlling external
    payment rails that the deploying party does not operate; and
    preventing harmful actions that are within a correctly specified
    authority. A valid policy can still authorize a bad decision.</t>
  </section>
</section>

<section anchor="terms" numbered="true">
  <name>Conventions and Terminology</name>
  <t>The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
  "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
  "OPTIONAL" in this document are to be interpreted as described in
  BCP&nbsp;14 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and
  only when, they appear in all capitals, as shown here. This is an
  Informational document; requirement keywords describe the profile, not an
  Internet standard.</t>
  <dl newline="false" spacing="normal">
    <dt>Candidate Act:</dt>
    <dd>An operation proposed by an agent, orchestrator or tool that has no
    external effect until released by a Finality Sink.</dd>
    <dt>Source-Authority Envelope (SAE):</dt>
    <dd>A protected record of what a principal actually authorized
    (purpose, permitted targets, recipients, amounts, scopes, phases,
    expiry, policy epoch). It is held outside any model-editable context.
    Prompt content, retrieved documents and model assertions cannot create
    or enlarge it.</dd>
    <dt>Candidate Act Descriptor (CAD):</dt>
    <dd>A canonical representation of the operation as it will actually be
    committed: resolved target resources, effective identity, destination,
    amounts, data scope, configuration delta and phase.</dd>
    <dt>Execution Handle:</dt>
    <dd>A short-lived, non-bearer authorization object bound to one CAD
    digest, one Finality Sink, one phase, the SAE, and the policy and
    revocation epochs. Possession alone does not authorize the effect.</dd>
    <dt>Finality Sink:</dt>
    <dd>The component that makes an effect real and owns the effect-causing
    credential or commit right (an export gateway, payment gateway,
    configuration admission controller, deployment controller, secret
    broker, response-release gate).</dd>
    <dt>Protected Enforcement Domain (PED):</dt>
    <dd>The execution environment in which the validator, consume state and
    sink signing keys reside, isolated from the agent runtime.</dd>
    <dt>Destination Receipt:</dt>
    <dd>An authenticated statement originated by the destination system of
    record (or an independent protected observer) of what was committed.
    A receipt is evidence; it is never usable as a handle.</dd>
    <dt>Continuation Authority:</dt>
    <dd>Authority for a later phase that is granted only on presentation of
    a valid receipt for the preceding phase.</dd>
    <dt>Covered Route:</dt>
    <dd>A path capable of producing a protected effect that is mediated by a
    Finality Sink, disabled, or restricted. Every other path is uncovered
    and MUST be declared as such.</dd>
    <dt>INDETERMINATE:</dt>
    <dd>The outcome classification used when the sink cannot establish
    whether an effect was committed (for example, timeout after
    submission).</dd>
  </dl>
  <t>Industry-standard equivalents, for readers who use other vocabulary:</t>
  <table>
    <thead><tr><th>Term here</th><th>Closest established term</th><th>Difference</th></tr></thead>
    <tbody>
      <tr><td>Finality Sink</td><td>Policy Enforcement Point at the resource; commit gate; reference monitor</td><td>Placed at the point where the effect becomes real, and owns the effect credential</td></tr>
      <tr><td>Execution Handle</td><td>Sender-constrained, fine-grained access token; capability</td><td>Bound to the digest of the exact act and to one sink, consumed with the commit</td></tr>
      <tr><td>SAE</td><td>Authorization details; mandate; change ticket</td><td>Held outside model-editable context; compared with the reconstructed act</td></tr>
      <tr><td>Destination Receipt</td><td>Transaction status, acknowledgement, signed log entry</td><td>Originated by the destination; gates continuation and retry</td></tr>
      <tr><td>INDETERMINATE</td><td>Unknown commit outcome; in-doubt transaction</td><td>Blocks retry until reconciled with the system of record</td></tr>
    </tbody>
  </table>
</section>

<section anchor="threat" numbered="true">
  <name>Adversary Model and Trust Assumptions</name>
  <section numbered="true">
    <name>Attacker Capabilities</name>
    <t>The attacker MAY: place instructions in retrieved documents, tickets,
    web pages, telemetry or tool outputs; mislead or fully control the
    agent's reasoning; cause the agent to invoke any tool the agent can
    reach with any syntactically valid arguments; substitute tool responses
    returned to the agent; replay or reorder requests; exploit broad
    credentials the agent already holds; cause timeouts and induce
    retries; delegate to other agents; and modify mutable business records
    the agent can write (for example, a beneficiary master record).</t>
  </section>
  <section numbered="true">
    <name>Assumptions</name>
    <dl newline="false">
      <dt>A1:</dt><dd>The Finality Sink and the validator are not both
      compromised. If both are, the controls in this document may fail.</dd>
      <dt>A2:</dt><dd>The SAE is created through a channel the agent cannot
      write (an authenticated human approval, a mandate store, a change
      system).</dd>
      <dt>A3:</dt><dd>Signature and hash primitives are not broken, and keys
      are held in the PED.</dd>
      <dt>A4:</dt><dd>The consume store is linearizable for each handle.</dd>
      <dt>A5:</dt><dd>The route inventory is complete for the routes
      declared as covered. Correctness of the inventory is an operational
      property that this document cannot prove.</dd>
      <dt>A6:</dt><dd>Policy correctness is enforced, not verified: the
      architecture enforces the SAE and policy as written.</dd>
    </dl>
  </section>
  <section numbered="true">
    <name>Engineering Invariants</name>
    <t>The following are engineering targets, not proved guarantees:</t>
    <dl newline="false">
      <dt>I1 No agent-derived authority:</dt><dd>No content reaching the model
      can create or enlarge an SAE.</dd>
      <dt>I2 Exact act:</dt><dd>A handle authorizes only the CAD digest it
      binds, at the sink it names, in the phase it names.</dd>
      <dt>I3 Same bytes:</dt><dd>The sink verifies the digest of the bytes it
      commits, not a digest supplied by the caller.</dd>
      <dt>I4 Single consumption:</dt><dd>Each handle produces at most one
      committed effect (or, for envelope handles, at most the declared count
      and aggregate ceiling).</dd>
      <dt>I5 No unobserved alternate effect:</dt><dd>Every known
      act-equivalent route is covered or declared uncovered.</dd>
      <dt>I6 No blind continuation:</dt><dd>A later consequential phase
      requires a valid receipt for the preceding phase.</dd>
      <dt>I7 No blind retry:</dt><dd>An INDETERMINATE outcome is reconciled
      before any retry of the same intent.</dd>
      <dt>I8 Attenuation:</dt><dd>Delegated authority is a subset of the
      delegator's authority.</dd>
    </dl>
  </section>
  <section numbered="true">
    <name>Informal Bound</name>
    <t>Under A1-A6, the probability that an unauthorized protected effect
    completes on a covered route is bounded informally by</t>
    <artwork type="ascii-art"><![CDATA[
  Pr[unauthorized effect on covered route]
      <= e_hash + e_sig + e_store + e_path + e_recon
]]></artwork>
    <t>where e_hash and e_sig are cryptographic failure probabilities,
    e_store is the probability of a consume-store linearizability failure,
    e_path the probability that a route believed covered is not, and
    e_recon the probability that the CAD reconstruction omits a load-bearing
    field. The last two terms dominate in practice and are the targets of
    the evaluation plan.</t>
  </section>
</section>

<section anchor="existing" numbered="true">
  <name>Relationship to Existing Controls and Guidance</name>
  <t>Resource-level policy enforcement is long established. NIST SP 800-207
  places a Policy Decision Point and Policy Enforcement Point in front of
  enterprise resources <xref target="NIST-SP-800-207"/>; NIST SP 800-162
  describes attribute-based decisions over subject, object, action and
  environment <xref target="NIST-SP-800-162"/>. Current agent-security
  guidance recommends scoped access, least privilege and runtime controls
  <xref target="OWASP-AGENTIC"/> <xref target="OWASP-CHEATSHEET"/>
  <xref target="CISA-AGENTIC"/>. This document builds on those
  mechanisms. The question a PEP typically answers is "may this subject
  access this resource with this action?" The additional question
  addressed here is "may this exact, reconstructed effect complete now,
  on this route, in this phase, given the destination evidence for
  earlier phases?"</t>
  <table>
    <thead><tr><th>Existing control</th><th>What it already prevents</th><th>Residual gap addressed here</th></tr></thead>
    <tbody>
      <tr><td>Prompt filtering / model guardrails</td><td>Many malicious instructions and disallowed outputs</td><td>Accepted injection or legitimate-looking misuse still reaches an authorized tool</td></tr>
      <tr><td>IAM, OAuth, RBAC/ABAC</td><td>Calls without valid identity, role, scope or policy</td><td>A broad valid token may not encode beneficiary, dataset, fan-out or side effect</td></tr>
      <tr><td>API gateway / agent firewall</td><td>Unapproved endpoints, syntax, quota, contextual policy</td><td>Hosted tools, direct credentials, queues or secondary APIs bypass an app-side gate</td></tr>
      <tr><td>DLP and egress control</td><td>Inspected, classified data movement</td><td>Encrypted or transformed payload; allowed destination with wrong purpose; unmanaged egress</td></tr>
      <tr><td>Payment risk and approval</td><td>Abnormal payments, limits, mandates</td><td>Approval-to-capture divergence; duplicate phase replay; connector drift</td></tr>
      <tr><td>SIEM / EDR / audit</td><td>Anomalies, investigation</td><td>Evidence after an irreversible effect; logs from compromised components</td></tr>
      <tr><td>Zero Trust resource PEP</td><td>Resource access with contextual policy</td><td>A complete effect-aware PEP may already implement much of this; added value is phase continuity and route closure</td></tr>
      <tr><td>OT interlocks / safety PLC</td><td>Unsafe physical states</td><td>A semantically wrong change that stays within safety limits</td></tr>
    </tbody>
  </table>
  <t>Where a deployment's PEP already reconstructs the effect, mediates all
  routes, binds phases and verifies outcomes, the incremental benefit of
  this profile may be marginal, and the profile serves mainly as a common
  representation for interoperability.</t>
</section>

<section anchor="arch" numbered="true">
  <name>Architecture Overview</name>
  <figure>
    <name>Trust-and-Effect Chain</name>
    <artwork type="ascii-art"><![CDATA[
 +-----------+ proposal +-------------+  CAD  +-------------+
 | AI agent  |--------->| Candidate   |------>| Validator   |
 | (no       |          | Act (NON_   |       | (PED): SAE, |
 | authority)|          | EFFECTIVE)  |       | epochs      |
 +-----------+          +-------------+       +------+------+
                                                     | Execution
                                                     v Handle
 +-------------+ receipt +---------------+ commit +----------+
 | Continua-   |<--------| Finality Sink |------->| Effect   |
 | tion / next |         | (owns effect  |        | (system  |
 | phase/retry |         | credential;   |        | of       |
 +-------------+         | recomputes    |        | record)  |
                         | CAD; consumes)|        +----------+
                         +---------------+
 Other routes to the same effect: covered by a sink, disabled,
 restricted, or declared UNCOVERED.
]]></artwork>
  </figure>
  <t>Minimum components: (1) an SAE held outside model-editable context;
  (2) a CAD resolved from actual arguments, target resources, effective
  identities and state; (3) an independent validator checking scope,
  policy and revocation epochs and protected counters; (4) a PED owning
  the effect-causing credential or commit right; (5) Finality Sinks on all
  covered routes; (6) destination-originated or observer-originated
  receipts; (7) HOLD and reconcile handling for unresolved effects.</t>
</section>

<section anchor="objects" numbered="true">
  <name>Data Objects</name>
  <t>Objects are shown as JSON for readability. Implementations MUST use a
  deterministic canonical form for digests: JSON Canonicalization Scheme
  <xref target="RFC8785"/> or deterministic CBOR <xref target="RFC8949"/>.
  Digests use SHA-256 unless a profile states otherwise. Field names are
  illustrative; no registry is requested by this document.</t>

  <section numbered="true">
    <name>Source-Authority Envelope</name>
    <sourcecode type="json"><![CDATA[
{
  "sae_id": "sae-7f3a",
  "principal": "user:finance-approver-12",
  "created_via": "approval-ui/authenticated",
  "purpose": "pay-invoice",
  "consequence_class": "FINANCIAL",
  "constraints": {
    "beneficiary_legal_entity": "Supplier A Ltd",
    "beneficiary_account_digest": "sha256:9c1e...",
    "invoice_ref": "INV-2026-0441",
    "currency": "EUR",
    "max_amount_minor": 1250000,
    "rails": ["SEPA-CT"],
    "phases": ["INITIATE"]
  },
  "policy_epoch": 418,
  "revocation_epoch": 77,
  "not_after": "2026-11-18T12:00:00Z",
  "sig": "<PED signature over canonical form>"
}
]]></sourcecode>
  </section>

  <section numbered="true">
    <name>Candidate Act Descriptor</name>
    <t>The CAD digest MUST cover every load-bearing field, that is, every
    field whose change would change the consequence. Fields that do not
    change the consequence (trace identifiers, formatting) SHOULD be
    excluded so that equivalent operations produce equal digests. Missing
    load-bearing information MUST be represented as UNKNOWN, never as an
    empty or default value, and UNKNOWN in a load-bearing field MUST cause
    HOLD or DENY.</t>
    <sourcecode type="json"><![CDATA[
{
  "cad_version": 1,
  "act_id": "act-55d1",
  "intent_id": "int-0441",
  "sae_id": "sae-7f3a",
  "act_type": "PAYMENT_INITIATE",
  "sink_id": "sink:treasury-gw-eu-1",
  "effective_identity": "svc:treasury-gw",
  "resolved": {
    "beneficiary_account_digest": "sha256:9c1e...",
    "beneficiary_name": "Supplier A Ltd",
    "amount_minor": 1250000,
    "currency": "EUR",
    "rail": "SEPA-CT",
    "fees_minor": 0
  },
  "phase": "INITIATE",
  "payload_digest": "sha256:<digest of final wire message>",
  "transitive_effects": []
}
]]></sourcecode>
  </section>

  <section numbered="true">
    <name>Execution Handle</name>
    <sourcecode type="json"><![CDATA[
{
  "eh_id": "eh-a90c",
  "cad_digest": "sha256:<digest of canonical CAD>",
  "sink_id": "sink:treasury-gw-eu-1",
  "phase": "INITIATE",
  "sae_id": "sae-7f3a",
  "policy_epoch": 418,
  "revocation_epoch": 77,
  "reuse_policy": "SINGLE_USE",
  "parent_eh": null,
  "delegation_depth": 0,
  "nbf": "2026-11-17T18:00:00Z",
  "exp": "2026-11-17T18:02:00Z",
  "holder_key_thumbprint": "<optional proof-of-possession key>",
  "sig": "<validator signature in PED>"
}
]]></sourcecode>
    <t>The handle is non-bearer in the sense that the sink authorizes only
    the act whose recomputed digest equals cad_digest; a holder that
    presents the handle with a different act obtains nothing. Holder key
    binding (for example DPoP <xref target="RFC9449"/>) MAY be added as an
    additional layer, but key binding without act and sink binding is a
    token-theft mitigation, not an exact-act control.</t>
  </section>

  <section numbered="true">
    <name>Destination Receipt</name>
    <sourcecode type="json"><![CDATA[
{
  "receipt_id": "rcpt-31b8",
  "act_id": "act-55d1",
  "eh_id": "eh-a90c",
  "cad_digest": "sha256:...",
  "sink_id": "sink:treasury-gw-eu-1",
  "origin": "system-of-record:psp-eu",
  "phase": "INITIATE",
  "outcome": "ACCEPTED_FOR_PROCESSING",
  "committed_state_digest": "sha256:...",
  "sor_reference": "PSP-TX-88123",
  "counter": 90211,
  "observed_at": "2026-11-17T18:00:41Z",
  "sig": "<signature by destination or protected observer>"
}
]]></sourcecode>
    <t>The outcome field MUST distinguish states that are legally or
    technically distinct (for payments: ACCEPTED_FOR_PROCESSING,
    AUTHORIZED, CAPTURED, SETTLED, REJECTED; for configuration: STAGED,
    ACTIVATED, READBACK_MATCHED, READBACK_MISMATCH). A transport-level
    success or agent text is not a receipt. Receipt authenticity is
    necessary but not sufficient: the origin must itself be trustworthy and
    must observe the consequential effect.</t>
  </section>
</section>

<section anchor="toctou" numbered="true">
  <name>Exact-Effect Binding at the Commit Point</name>
  <t>A check performed on a request and an effect committed from different
  bytes leaves a time-of-check to time-of-use window. Connectors, retry
  workers, template expansion, beneficiary lookup and controller rendering
  can all change the operation after approval. The sink therefore MUST:</t>
  <ol>
    <li>construct the final committable message (wire payload, SQL
    statement and parameters, rendered configuration, deployment manifest)
    before verification;</li>
    <li>recompute the CAD from that final message and from current resolved
    state (for example, the beneficiary account the connector will actually
    use), never trusting a CAD or digest supplied by the caller;</li>
    <li>verify the handle signature, expiry, sink, phase and cad_digest
    against the recomputed CAD (VERIFY_STATIC);</li>
    <li>inside the same atomic section as consume and commit, re-read the
    current policy epoch, revocation epoch and protected counters
    (VERIFY_CURRENT);</li>
    <li>commit exactly the verified bytes, with no further transformation
    after verification.</li>
  </ol>
  <figure>
    <name>Sink Procedure (Pseudocode)</name>
    <sourcecode type="pseudocode"><![CDATA[
function sink_commit(handle, request):
    msg  = build_final_message(request)   # bytes to be committed
    cad  = reconstruct_cad(msg, resolve_live_state(msg))
    if cad.has_unknown_load_bearing():
        return HOLD("EF-UNKNOWN-FIELD")
    if not verify_static(handle, digest(cad), SINK_ID):
        return DENY("EF-STATIC")
    begin_atomic(handle.eh_id):
        if not verify_current(handle, epochs(), counters()):
            abort_atomic(); return DENY("EF-STALE")
        if consume_state(handle.eh_id) != UNUSED:
            abort_atomic(); return DENY("EF-REPLAY")
        write_consume(handle.eh_id, PENDING, digest(cad))   # durable
        result = commit(msg)                     # the effect
        if result.known:
            write_consume(handle.eh_id, COMMITTED, result.ref)
        else:
            write_consume(handle.eh_id, INDETERMINATE, null)
    end_atomic()
    return emit_receipt(handle, cad, result)
]]></sourcecode>
  </figure>
  <t>When the destination is external and cannot participate in the atomic
  section (for example, a third-party payment API), the sink MUST record
  PENDING durably before submission and MUST use a destination
  idempotency key derived from act_id, so that a later reconciliation can
  establish whether the submission took effect.</t>
</section>

<section anchor="state-machine" numbered="true">
  <name>Candidate Act State Machine</name>
  <figure>
    <name>States and Transitions</name>
    <artwork type="ascii-art"><![CDATA[
  NON_EFFECTIVE --validate ok--> AUTHORIZED --commit ok--> COMMITTED
     |     |                       ^  |   |
     |     +--needs approval--> HELD  |   +--no ack--> INDETERMINATE
     |                                |                     |
     |                   expiry --> EXPIRED            reconcile
     |                                                      |
     +--mismatch / UNKNOWN / stale --> DENIED               v
                                          RECONCILED_COMMITTED or
                                          RECONCILED_NOT_COMMITTED
]]></artwork>
  </figure>
  <table>
    <thead><tr><th>Transition</th><th>Triggered only by</th><th>Condition</th></tr></thead>
    <tbody>
      <tr><td>NON_EFFECTIVE to AUTHORIZED</td><td>Validator</td><td>Reconstructed CAD within SAE; epochs current; handle issued</td></tr>
      <tr><td>NON_EFFECTIVE to HELD</td><td>Validator</td><td>Policy requires approval, or scope exceeds SAE but is approvable</td></tr>
      <tr><td>HELD to AUTHORIZED</td><td>Validator, after authenticated approval</td><td>Approval bound to the same CAD digest (what is displayed is what is executed)</td></tr>
      <tr><td>any to DENIED</td><td>Validator or sink</td><td>Mismatch, UNKNOWN load-bearing field, revoked or stale authority</td></tr>
      <tr><td>AUTHORIZED to COMMITTED</td><td>Sink</td><td>Atomic consume and commit with known outcome</td></tr>
      <tr><td>AUTHORIZED to INDETERMINATE</td><td>Sink</td><td>Commit attempted; outcome unknown</td></tr>
      <tr><td>INDETERMINATE to RECONCILED_*</td><td>Reconciler using system of record</td><td>Authoritative status obtained</td></tr>
      <tr><td>AUTHORIZED to EXPIRED</td><td>Sink or validator</td><td>Handle past exp without consumption</td></tr>
    </tbody>
  </table>
  <t>The agent triggers no transition. A new attempt for the same intent_id
  while any act for that intent is AUTHORIZED, HELD or INDETERMINATE MUST
  be refused (one live authority per intent). Compensation for a committed
  effect is a new Candidate Act with a new handle; it is never an undo of a
  consumed handle. The state machine is small enough to be checked with
  model-checking tools such as TLA+; doing so is part of the evaluation
  plan.</t>
</section>

<section anchor="consume" numbered="true">
  <name>Atomic Consumption, Idempotency and Crash Recovery</name>
  <ul>
    <li>Consume state is keyed per handle (compare-and-swap on a per-handle
    record), not a global lock, so that unrelated acts do not serialize.</li>
    <li>For sharded stores, the handle MUST be routed to the shard that owns
    its eh_id; cross-shard consumption of one handle is not permitted.</li>
    <li>The PENDING record MUST be durable before the effect is attempted
    (write-ahead). After a crash, a handle found in PENDING is treated as
    INDETERMINATE, never as UNUSED.</li>
    <li>On failover, the new primary MUST NOT accept consumption for any
    handle until its consume store is at least as recent as the last
    acknowledged write of the failed primary; otherwise it MUST HOLD.</li>
    <li>Exactly-once is claimed only relative to the sink: at most one
    commit attempt per handle. Exactly-once at an external destination
    additionally requires destination idempotency on act_id.</li>
    <li>If the consume store is unavailable, the sink MUST fail closed for
    the protected consequence class (HOLD), subject to the safety rules in
    <xref target="scenario-c"/>.</li>
  </ul>
</section>

<section anchor="sink-trust" numbered="true">
  <name>Trust in the Sink and the Validator</name>
  <ul>
    <li>Validator signing keys and sink receipt keys SHOULD reside in an HSM
    or trusted execution environment within the PED, and MUST NOT be
    reachable from the agent runtime.</li>
    <li>Before accepting handles for a consequence class, a sink SHOULD
    present remote attestation evidence of its software identity and
    configuration to the validator, using the RATS architecture
    <xref target="RFC9334"/>. The validator SHOULD issue handles only to
    sinks whose attestation results are current.</li>
    <li>Validator and sink SHOULD be administered by separate roles, so that
    assumption A1 is not defeated by one administrative compromise.</li>
    <li>Key rotation MUST NOT invalidate the ability to verify receipts
    already issued; receipts carry key identifiers.</li>
    <li>On suspected key compromise, the revocation epoch is advanced,
    outstanding handles signed by the affected key are rejected, and
    receipts signed during the suspected window are treated as
    unverified pending reconciliation.</li>
    <li>For high-consequence classes, multi-party validation (for example,
    two validators in different trust domains) MAY be required.</li>
  </ul>
</section>

<section anchor="delegation" numbered="true">
  <name>Delegation and Attenuation in Multi-Agent Systems</name>
  <t>When agent A delegates a sub-task to agent B, B does not receive A's
  credentials. B receives, at most, the ability to request handles under a
  derived SAE. The following rules apply:</t>
  <ol>
    <li>A derived SAE MUST be a subset of its parent in every constraint
    dimension (targets, recipients, amounts, data scope, phases, expiry).
    The validator, not the delegating agent, computes and signs the derived
    SAE.</li>
    <li>Aggregate ceilings are shared: the sum of effects under all
    derived SAEs MUST NOT exceed the parent ceiling. The validator tracks
    this in protected counters.</li>
    <li>Handles record parent_eh and delegation_depth; a maximum depth is
    set per consequence class.</li>
    <li>Revoking a parent SAE revokes all derived SAEs.</li>
    <li>A delegated agent's statement that a sub-task succeeded is not a
    receipt; continuation depends on destination receipts.</li>
  </ol>
  <t>This is the attenuation property of capability systems applied to
  agent delegation; it prevents a sub-agent from accumulating authority
  that no human principal granted.</t>
</section>

<section anchor="protocol-placement" numbered="true">
  <name>Placement in Tool-Call and Agent-to-Agent Protocols</name>
  <section numbered="true">
    <name>Tool Calls</name>
    <t>For a tool protocol such as the Model Context Protocol
    <xref target="MCP"/>, the tool server that holds the effect credential
    is the natural Finality Sink. A "tools/call" request is a Candidate Act.
    The tool server reconstructs the CAD from the arguments it will
    actually use and from live state, and commits only on a valid handle.
    Tools that only read non-sensitive data need not participate. Hosted or
    provider-side tools that the deploying party cannot instrument are
    uncovered routes unless the destination enforces.</t>
    <t>An illustrative HTTP carriage, not a registered scheme:</t>
    <sourcecode><![CDATA[
POST /v1/payments HTTP/1.1
Host: treasury-gw.example
Authorization: EFHandle eh=<base64url(handle)>
Idempotency-Key: act-55d1
Content-Type: application/json
]]></sourcecode>
    <t>Failures SHOULD be returned as problem details <xref target="RFC9457"/>
    with an explicit reason (UNKNOWN field, stale epoch, replay, mismatch).</t>
  </section>
  <section numbered="true">
    <name>Agent-to-Agent Calls</name>
    <t>A request from one agent to another is not itself an effect. The
    receiving agent's eventual effects pass through their own sinks under a
    derived SAE (<xref target="delegation"/>). Agent-to-agent messages
    SHOULD carry the parent SAE identifier and intent_id so that one live
    authority per intent can be enforced across agents.</t>
  </section>
  <section anchor="ietf-relation" numbered="true">
    <name>Relationship to OAuth Mechanisms</name>
    <table>
      <thead><tr><th>Mechanism</th><th>What it provides</th><th>What this profile adds</th></tr></thead>
      <tbody>
        <tr><td>OAuth 2.0 <xref target="RFC6749"/></td><td>Delegated access tokens with scopes</td><td>Scopes do not bind a resolved act; the handle binds a CAD digest and a sink</td></tr>
        <tr><td>Rich Authorization Requests <xref target="RFC9396"/></td><td>Structured authorization_details describing an intended act</td><td>RAR can describe the act; it does not by itself require the resource to recompute the act from final bytes, consume authority with the commit, or gate continuation on a destination receipt. RAR authorization_details are a suitable carrier for SAE constraints.</td></tr>
        <tr><td>Token Exchange <xref target="RFC8693"/></td><td>Exchange of one token for another, including delegation and impersonation semantics</td><td>Attenuation rules and shared aggregate ceilings across derived authorities</td></tr>
        <tr><td>DPoP <xref target="RFC9449"/></td><td>Sender-constrained tokens</td><td>Binding to the act and sink, not only to a key</td></tr>
        <tr><td>Transaction Tokens <xref target="TXN-TOKENS"/></td><td>Propagation of immutable context through a call chain inside a trust domain</td><td>Commit-time verification, consumption and receipt-gated continuation at the effect point</td></tr>
      </tbody>
    </table>
    <t>A practical deployment can therefore carry SAE constraints in RAR,
    derive sub-agent authority with Token Exchange, constrain the holder
    with DPoP, propagate context with Transaction Tokens, and add the
    act-bound handle, sink recomputation, consumption and receipts
    described here.</t>
  </section>
</section>

<section anchor="revocation" numbered="true">
  <name>Revocation and Freshness</name>
  <t>The validator maintains a monotonic policy epoch and revocation epoch
  per consequence class. Sinks hold a protected local copy. The maximum
  time during which a revoked authority could still be honoured on a
  covered route is bounded by</t>
  <artwork type="ascii-art"><![CDATA[
  W_stale <= min( handle_lifetime , D_prop + d_sink ) + T_hot

  D_prop : maximum time for an epoch change to reach the sink
  d_sink : maximum sink cache age before mandatory refresh
  T_hot  : sink verify-to-commit time
]]></artwork>
  <t>Each consequence class SHOULD declare a target W_stale. If the sink
  cannot confirm its epoch copy is within d_sink, it MUST HOLD.</t>
  <t>Offline leases. Where connectivity to the validator is intermittent
  (edge, OT, satellite), a protected edge controller MAY hold a signed,
  time-bounded lease for a pre-approved operation set. The lease declares
  T_off_max (maximum offline exposure), a monotonic counter that cannot be
  rolled back, and the safe state on expiry. Revocation lag during offline
  operation is bounded by T_off_max and MUST be accepted explicitly in the
  safety case.</t>
</section>

<section anchor="scenario-a" numbered="true">
  <name>Scenario A: AI-Driven Enterprise Data Exfiltration</name>
  <t>Legitimate work: an agent summarizes supplier contracts and sends
  approved reports to a known partner. Attack: indirect prompt injection
  in a supplier PDF directs the agent to query an expanded confidential
  dataset and upload it to an attacker-controlled bucket.</t>
  <table>
    <thead><tr><th>Stage</th><th>What may happen</th><th>Consequence to prevent</th></tr></thead>
    <tbody>
      <tr><td>Ingest</td><td>Agent treats document text as instruction</td><td>Untrusted content as instruction source</td></tr>
      <tr><td>Plan</td><td>Broad query and export endpoint chosen</td><td>Privilege used outside purpose</td></tr>
      <tr><td>Query</td><td>Authorized read identity retrieves large set</td><td>Collection beyond business scope</td></tr>
      <tr><td>Export</td><td>Upload to unapproved domain or bucket</td><td>Irreversible disclosure</td></tr>
      <tr><td>Continue</td><td>Agent reports success, starts more exports</td><td>Cascading exfiltration</td></tr>
    </tbody>
  </table>
  <t>Profile: the SAE records the purpose, permitted classifications,
  recipients, columns or row predicates, volume ceiling and expiry. The
  export service or egress proxy is the sink and owns the upload
  credential; the agent never holds it. The sink reconstructs the CAD from
  the effective query plan or result-set classification, row count,
  tenant, encryption target, resolved destination account and endpoint,
  and the content hash of the final object or manifest. Mismatched
  recipient, classification or scope yields HOLD or DENY. The destination
  or export gateway emits the receipt; further exports require it, and
  the volume ceiling is enforced cumulatively across acts (closing
  fragmented exfiltration).</t>
  <t>Placement in existing estates: storage gateway or bucket policy
  requiring a scoped export authorization; a controlled export procedure
  for SQL, with trigger and definer-rights effects included in the CAD;
  the tool server's broad service credential replaced with a
  non-effectuating request interface; DLP and TLS inspection retained.</t>
  <t>Residual: an approved destination later compromised; inaccurate
  classification; scheduled jobs outside the gateway; leakage in the
  model's own response (<xref target="output-channel"/>); onward forwarding
  by an allowed recipient; side channels.</t>
</section>

<section anchor="scenario-b" numbered="true">
  <name>Scenario B: AI-Initiated Payment Diversion</name>
  <t>A treasury agent is asked to pay a valid invoice for Supplier A. An
  attacker manipulates a vendor instruction, purchase-order field, stored
  beneficiary record or connector response so the submitted transaction
  pays another account. The request is well formed and sent under a
  legitimate credential.</t>
  <t>Profile: the SAE binds payer, legal beneficiary and account digest,
  invoice reference, currency, amount, FX conditions, rail, phase and
  expiry. The payment gateway is the sink and holds the transaction
  credential. It reconstructs the CAD from the beneficiary and routing data
  the connector will actually submit (not the agent's displayed
  confirmation), including fees, FX and split payments. Exposure is
  reserved atomically per invoice and act_id; split captures that together
  exceed the approved total are refused by the cumulative counter. Each
  phase (initiate, authorize, capture, settle) is a separate act with its
  own handle; continuation requires a receipt identifying the actual
  accepted or posted state from the PSP or bank. A timeout becomes
  INDETERMINATE and is reconciled using the idempotency key before any
  retry. Payment rails and participant rules remain authoritative; the
  profile does not control external settlement.</t>
  <t>Residual: beneficiary master data that is wrong at approval time;
  partner APIs that do not expose transaction identity or idempotency;
  settled payments that are costly to reverse. The profile aims to prevent
  unauthorized phases, not to promise rollback.</t>
</section>

<section anchor="scenario-c" numbered="true">
  <name>Scenario C: Critical-Infrastructure Configuration Manipulation</name>
  <t>An operations agent may propose changes to a network slice, an
  industrial gateway or an OT management system. Injected telemetry or
  ticket text induces a change outside the approved envelope: a policy
  rule, route redirection, slice resource change, security-control
  disablement or unsafe maintenance command.</t>
  <t>Profile: the SAE binds the change ticket, target nodes or slices,
  configuration delta, maintenance window, topology assumptions and
  permitted failure mode. Changes are staged as non-effective. The sink is
  the controller admission point, configuration activation or hardware-rule
  installation step. The CAD is reconstructed from the actual target set,
  dependent resources, transitive controller effects and generated
  hardware or network rules, compared with live desired and observed state.
  The receipt is an acknowledged installed-state version or independent
  readback; fleet-scale continuation requires bounded verified outcomes
  from an initial subset where the effect is divisible.</t>
  <t>Placement: slice or service-management configuration activation and
  security-policy commit; non-real-time control policy and approved
  application deployment in O-RAN management; remote engineering
  commands and recipe or setpoint changes in OT; Kubernetes admission and
  firewall or route installation in cloud and edge. Per-frame, per-packet,
  scheduler and near-real-time control loops are not suitable boundaries
  without measured proof.</t>
  <t>Safety: this profile complements safety-certified interlocks and MUST
  NOT replace them or delay a protective action <xref target="NIST-SP-800-82"/>.
  "Fail closed" is not universally safe. When validation is unavailable,
  the system enters the safe state selected by its operational safety case.
  Vendor maintenance ports, local panels and alternate automation paths
  that retain unrestricted authority are uncovered routes.</t>
</section>

<section anchor="scenario-d" numbered="true">
  <name>Scenario D: Code-Executing and Computer-Use Agents</name>
  <t>Coding and computer-use agents run shell commands, edit repositories,
  push commits, trigger CI/CD pipelines, and request secrets. Injected
  content in an issue, dependency README or web page can direct the agent
  to add a malicious dependency, exfiltrate environment variables, push to
  a protected branch, or deploy to production.</t>
  <t>Consequential boundaries in this setting are few and identifiable:</t>
  <table>
    <thead><tr><th>Boundary</th><th>Sink</th><th>CAD fields</th></tr></thead>
    <tbody>
      <tr><td>Secret release</td><td>Secret broker</td><td>Secret identifier, requesting workload, destination process or environment, purpose, lifetime</td></tr>
      <tr><td>Push / merge to protected branch</td><td>Repository server pre-receive or merge gate</td><td>Repository, branch, commit digest set, changed paths, dependency manifest delta</td></tr>
      <tr><td>Pipeline or deployment</td><td>Deployment controller</td><td>Artifact digest, target environment, manifest digest, rollout scope</td></tr>
      <tr><td>Network egress from the sandbox</td><td>Egress proxy</td><td>Resolved destination, payload digest, volume</td></tr>
      <tr><td>Package publication</td><td>Registry publish gate</td><td>Package name, version, artifact digest</td></tr>
    </tbody>
  </table>
  <t>Commands confined to an isolated workspace without credentials or
  egress are not consequential and need no handle; this keeps the interactive
  loop fast. The agent's sandbox SHOULD hold no long-lived credentials; the
  secret broker releases only act-bound, short-lived secrets into a
  destination it can identify. A CI job's report of success is not a
  receipt; the deployment controller's readback of the running artifact
  digest is.</t>
</section>

<section anchor="output-channel" numbered="true">
  <name>Output-Channel (Response Egress) Profile</name>
  <t>Data can leave through the agent's own response to a user or to
  another system, without any export tool. Where the response is delivered
  across a trust boundary (to an external user, a partner system, or a
  lower-classification tenant), the response-release point can be treated
  as a Finality Sink: the CAD records the recipient, the classifications of
  sources that contributed to the response (via retrieval provenance), and
  a digest of the response; the SAE states which classifications may be
  released to that recipient. Missing provenance is UNKNOWN, not "no
  sensitive sources". This profile limits disclosure scope; it does not
  detect paraphrase of protected content from sources whose provenance was
  never recorded, and it is not a content classifier.</t>
</section>

<section anchor="latency" numbered="true">
  <name>Latency, Throughput and Availability</name>
  <t>Design principle: move heavy validation away from high-rate inference,
  packet and actuator paths into rare consequential commit boundaries.</t>
  <artwork type="ascii-art"><![CDATA[
  T_finality = T_intercept + T_reconstruct + T_policy
             + T_protected_state + T_crypto + T_sink_commit
             + T_optional_remote
]]></artwork>
  <t>The quantity of interest is incremental p50/p95/p99/p99.9 delay
  relative to a comparable best-practice baseline, not an isolated
  signature-verification benchmark. Overhead-limiting strategies:
  precompile policy and target descriptors (cold path) so the sink performs
  deterministic verification (hot path); co-locate the validator with the
  sink; gate only consequence classes; bind a canonical manifest for
  batches instead of signing each row or packet, and refuse hidden
  additions after manifest approval; verify a signed lease once and then
  apply a lightweight local match until expiry; deliver receipts
  asynchronously except where they gate the next phase.</t>
  <t>Tail hazards include remote policy stores, HSM contention, cold
  starts, partitions, counter lock contention and shared receipt services.
  Under overload, the sink MUST NOT silently fall back to allow; bounded
  queues, partitioned counters, circuit-breaker HOLD and pre-authorized
  emergency paths are preferred. Monitor false-denial rate,
  non-effective dwell time, receipt age, reconciliation time and cache
  staleness. No specific millisecond figure is asserted by this
  document.</t>
</section>

<section anchor="legacy" numbered="true">
  <name>Legacy Integration</name>
  <table>
    <thead><tr><th>Available point</th><th>Approach</th><th>Coverage claim</th></tr></thead>
    <tbody>
      <tr><td>Native transaction hook</td><td>Plug in reconstruction and protected commit authorization</td><td>Strong if all equivalent routes use the hook</td></tr>
      <tr><td>Mature resource PEP</td><td>Extend with act and phase binding; reuse policy</td><td>May already satisfy much of the model</td></tr>
      <tr><td>API gateway or tool dispatcher</td><td>Hold the effect credential; inspect material arguments</td><td>Covered routes only</td></tr>
      <tr><td>Queue or workflow engine</td><td>Consumers verify handle before commit; outbox and idempotency</td><td>Requires all consumers to enforce</td></tr>
      <tr><td>Closed appliance</td><td>Management boundary, credential isolation, compensating controls</td><td>No native sink claim without proof</td></tr>
    </tbody>
  </table>
  <t>Migration in five stages: inventory every act-equivalent path
  (credentials, admin channels, scheduled jobs, delegated workers, queue
  consumers, maintenance interfaces); observe with shadow policy and CAD
  logging; constrain by removing broad credentials from agents and
  inserting one sink at one high-value boundary; enforce with consumption
  and HOLD/reconcile semantics; expand only after alternate-path closure
  is demonstrated.</t>
  <t>What cannot be promised: a shim in front of a service does not protect
  a direct admin API that bypasses it; a token in an agent runtime does not
  make an external rail secure; a signed receipt from a compromised
  destination does not prove the truthful effect; a non-bypassability
  statement is limited to the resources, versions and routes for which
  coverage has been demonstrated.</t>
</section>

<section anchor="bypass" numbered="true">
  <name>Bypass, Replay and Failure-Mode Analysis</name>
  <table>
    <thead><tr><th>Case</th><th>Rule</th><th>Residual</th></tr></thead>
    <tbody>
      <tr><td>Injection changes agent intent</td><td>Retrieved content never amends the SAE</td><td>In-scope harmful acts remain possible</td></tr>
      <tr><td>Direct credential bypass</td><td>Remove unmanaged credentials; sink rejects acts without handles</td><td>Unmanaged credential = uncovered route</td></tr>
      <tr><td>Destination changed after approval</td><td>Recompute CAD at commit from final bytes</td><td>Endpoint and beneficiary resolution must be controlled</td></tr>
      <tr><td>Stale policy or revocation</td><td>VERIFY_CURRENT inside atomic section</td><td>Offline lease window</td></tr>
      <tr><td>Replay or parallel race</td><td>Per-handle CAS; one live authority per intent</td><td>Failover consistency</td></tr>
      <tr><td>Lost acknowledgement</td><td>INDETERMINATE; reconcile before retry</td><td>Delayed finality at system of record</td></tr>
      <tr><td>Fragmentation</td><td>Cumulative ceilings across acts and delegates</td><td>Cross-principal decomposition</td></tr>
      <tr><td>Receipt presented as handle</td><td>Distinct object types; sink rejects</td><td>None expected</td></tr>
      <tr><td>Approval fatigue</td><td>Approval bound to CAD digest and shown in full</td><td>Human error</td></tr>
      <tr><td>Hardware offload mismatch</td><td>Readback of installed rules before continuation</td><td>Hardware API availability</td></tr>
      <tr><td>Validator unavailable</td><td>HOLD or predeclared lease; never silent bypass</td><td>Availability and safety trade-off</td></tr>
      <tr><td>Malicious validator or sink</td><td>Attestation, separation of roles, multi-party checks</td><td>Joint compromise outside guarantee</td></tr>
    </tbody>
  </table>
</section>

<section anchor="evaluation" numbered="true">
  <name>Falsifiable Evaluation Plan</name>
  <t>Test arms: A, a basic agent with ordinary tool access (measures the
  unprotected failure mode only); B, a strong conventional baseline with
  guardrails, scoped IAM, DLP and egress, a resource PEP, approvals,
  logging, idempotency and native safeguards; C, the B baseline plus this
  profile.</t>
  <t>Attack variants include: prompt injection; tool-output substitution;
  benign-looking wrong endpoint; scope escalation; multi-step delegation;
  stale authority after revocation; reordering; duplicate webhook; lost
  acknowledgement; expired handle; direct credential route; alternate API
  route; stored-procedure transitive effect; fragmented exfiltration;
  split payment capture; changed beneficiary master record; replayed
  configuration; controller reconciliation to stale state; validator
  outage; failover race; compromised receipt source; poisoned dependency
  in a coding task; secret request to an unidentified destination.</t>
  <table>
    <thead><tr><th>Measure</th><th>Acceptance condition</th></tr></thead>
    <tbody>
      <tr><td>Unauthorized commits at covered sinks</td><td>Target zero over the test corpus; publish coverage limits</td></tr>
      <tr><td>Alternate-path completions</td><td>Target zero on routes declared covered</td></tr>
      <tr><td>Legitimate task completion</td><td>No unacceptable false-denial regression</td></tr>
      <tr><td>Incremental tail latency</td><td>Within the declared SLO per consequence class</td></tr>
      <tr><td>Outcome classification</td><td>No silent retry after unresolved effect</td></tr>
      <tr><td>Revocation</td><td>Observed staleness within declared W_stale</td></tr>
      <tr><td>Safety</td><td>No missed safety-case requirement</td></tr>
    </tbody>
  </table>
  <t>Formal checks: model the state machine of <xref target="state-machine"/>
  and the consume rules of <xref target="consume"/> in a model checker and
  check I4 (single consumption) and I7 (no blind retry) under crash,
  failover and message reordering. An advantage exists only if arm C blocks
  cases arm B does not, at acceptable latency, cost and availability. A
  negligible difference against a complete effect-aware baseline is a
  valid result and should be reported.</t>
</section>

<section anchor="ietf-wg" numbered="true">
  <name>Relevance to IETF Work</name>
  <ul>
    <li>OAUTH: carriage of SAE constraints in authorization_details;
    act-bound, sink-bound handles as a token profile; attenuation rules
    for Token Exchange.</li>
    <li>WIMSE: workload identity for sinks, validators and agent
    runtimes; propagation of intent_id and SAE identifiers across
    workloads.</li>
    <li>RATS: attestation of sinks and validators before handle issuance.</li>
    <li>SCITT and COSE: transparent, verifiable destination receipts.</li>
    <li>GNAP: interaction-bound approval of a specific act.</li>
    <li>SECDISPATCH and DISPATCH: venue for agent effect-control work if
    none of the above fits.</li>
  </ul>
</section>

<section anchor="limits" numbered="true">
  <name>Limitations and Open Problems</name>
  <ul>
    <li>CAD semantic completeness: proving that a reconstruction captures
    every load-bearing field is hard for complex resources (SQL with
    triggers, controller reconciliation).</li>
    <li>Route completeness cannot be proved in general; it can only be
    demonstrated for declared routes.</li>
    <li>Cross-principal decomposition: a prohibited objective split into
    individually authorized acts by different principals.</li>
    <li>Authorized-but-harmful actions are not prevented.</li>
    <li>Non-preparable destinations (email, many external APIs) limit
    atomicity to the sink side.</li>
    <li>Wire formats and registries are illustrative.</li>
    <li>No performance figures are claimed; all must be measured.</li>
  </ul>
</section>

<section anchor="security" numbered="true">
  <name>Security Considerations</name>
  <t>This document is about security. The principal risks of deploying it
  are: (1) false confidence from declaring routes covered that are not;
  (2) incomplete CAD reconstruction that authorizes an act differing in an
  unmodelled field; (3) denial of service through deliberate HOLD-inducing
  requests or validator overload; (4) unsafe fail-closed behaviour in
  physical systems; (5) treating receipts from compromised origins as
  truth; (6) canonicalization disagreements between validator and sink
  implementations in different languages. Mitigations: publish declared
  coverage with each deployment; fail closed on UNKNOWN load-bearing
  fields and canonicalization failure; rate-limit Candidate Acts per
  principal and intent; define safe states per safety case; diversify
  receipt origins; maintain cross-implementation canonicalization test
  vectors including negative vectors (duplicate keys, Unicode
  normalization, numeric forms).</t>
</section>

<section anchor="privacy" numbered="true">
  <name>Privacy Considerations</name>
  <t>SAEs, CADs and receipts can contain personal data (beneficiary names,
  data subjects, recipients). Implementations SHOULD place digests rather
  than values in objects that leave the PED, retain receipts only for the
  period required by the consequence class, and avoid using receipt logs
  as a general activity record of users.</t>
</section>

<section anchor="iana" numbered="true">
  <name>IANA Considerations</name>
  <t>This document has no IANA actions. Identifiers shown are
  illustrative.</t>
</section>

</middle>

<back>
<references>
  <name>References</name>
  <references>
    <name>Normative References</name>
    <reference anchor="RFC2119" target="https://www.rfc-editor.org/info/rfc2119">
      <front><title>Key words for use in RFCs to Indicate Requirement Levels</title>
      <author initials="S." surname="Bradner" fullname="S. Bradner"/>
      <date year="1997" month="March"/></front>
      <seriesInfo name="BCP" value="14"/><seriesInfo name="RFC" value="2119"/>
      <seriesInfo name="DOI" value="10.17487/RFC2119"/>
    </reference>
    <reference anchor="RFC8174" target="https://www.rfc-editor.org/info/rfc8174">
      <front><title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title>
      <author initials="B." surname="Leiba" fullname="B. Leiba"/>
      <date year="2017" month="May"/></front>
      <seriesInfo name="BCP" value="14"/><seriesInfo name="RFC" value="8174"/>
      <seriesInfo name="DOI" value="10.17487/RFC8174"/>
    </reference>
  </references>
  <references>
    <name>Informative References</name>
    <reference anchor="RFC6749" target="https://www.rfc-editor.org/info/rfc6749">
      <front><title>The OAuth 2.0 Authorization Framework</title>
      <author initials="D." surname="Hardt" fullname="D. Hardt" role="editor"/>
      <date year="2012" month="October"/></front>
      <seriesInfo name="RFC" value="6749"/><seriesInfo name="DOI" value="10.17487/RFC6749"/>
    </reference>
    <reference anchor="RFC8693" target="https://www.rfc-editor.org/info/rfc8693">
      <front><title>OAuth 2.0 Token Exchange</title>
      <author initials="M." surname="Jones" fullname="M. Jones"/>
      <author initials="A." surname="Nadalin" fullname="A. Nadalin"/>
      <author initials="B." surname="Campbell" fullname="B. Campbell" role="editor"/>
      <author initials="J." surname="Bradley" fullname="J. Bradley"/>
      <author initials="C." surname="Mortimore" fullname="C. Mortimore"/>
      <date year="2020" month="January"/></front>
      <seriesInfo name="RFC" value="8693"/><seriesInfo name="DOI" value="10.17487/RFC8693"/>
    </reference>
    <reference anchor="RFC8785" target="https://www.rfc-editor.org/info/rfc8785">
      <front><title>JSON Canonicalization Scheme (JCS)</title>
      <author initials="A." surname="Rundgren" fullname="A. Rundgren"/>
      <author initials="B." surname="Jordan" fullname="B. Jordan"/>
      <author initials="S." surname="Erdtman" fullname="S. Erdtman"/>
      <date year="2020" month="June"/></front>
      <seriesInfo name="RFC" value="8785"/><seriesInfo name="DOI" value="10.17487/RFC8785"/>
    </reference>
    <reference anchor="RFC8949" target="https://www.rfc-editor.org/info/rfc8949">
      <front><title>Concise Binary Object Representation (CBOR)</title>
      <author initials="C." surname="Bormann" fullname="C. Bormann"/>
      <author initials="P." surname="Hoffman" fullname="P. Hoffman"/>
      <date year="2020" month="December"/></front>
      <seriesInfo name="STD" value="94"/><seriesInfo name="RFC" value="8949"/>
      <seriesInfo name="DOI" value="10.17487/RFC8949"/>
    </reference>
    <reference anchor="RFC9334" target="https://www.rfc-editor.org/info/rfc9334">
      <front><title>Remote ATtestation procedureS (RATS) Architecture</title>
      <author initials="H." surname="Birkholz" fullname="H. Birkholz"/>
      <author initials="D." surname="Thaler" fullname="D. Thaler"/>
      <author initials="M." surname="Richardson" fullname="M. Richardson"/>
      <author initials="N." surname="Smith" fullname="N. Smith"/>
      <author initials="W." surname="Pan" fullname="W. Pan"/>
      <date year="2023" month="January"/></front>
      <seriesInfo name="RFC" value="9334"/><seriesInfo name="DOI" value="10.17487/RFC9334"/>
    </reference>
    <reference anchor="RFC9396" target="https://www.rfc-editor.org/info/rfc9396">
      <front><title>OAuth 2.0 Rich Authorization Requests</title>
      <author initials="T." surname="Lodderstedt" fullname="T. Lodderstedt"/>
      <author initials="J." surname="Richer" fullname="J. Richer"/>
      <author initials="B." surname="Campbell" fullname="B. Campbell"/>
      <date year="2023" month="May"/></front>
      <seriesInfo name="RFC" value="9396"/><seriesInfo name="DOI" value="10.17487/RFC9396"/>
    </reference>
    <reference anchor="RFC9449" target="https://www.rfc-editor.org/info/rfc9449">
      <front><title>OAuth 2.0 Demonstrating Proof of Possession (DPoP)</title>
      <author initials="D." surname="Fett" fullname="D. Fett"/>
      <author initials="B." surname="Campbell" fullname="B. Campbell"/>
      <author initials="J." surname="Bradley" fullname="J. Bradley"/>
      <author initials="T." surname="Lodderstedt" fullname="T. Lodderstedt"/>
      <author initials="M." surname="Jones" fullname="M. Jones"/>
      <author initials="D." surname="Waite" fullname="D. Waite"/>
      <date year="2023" month="September"/></front>
      <seriesInfo name="RFC" value="9449"/><seriesInfo name="DOI" value="10.17487/RFC9449"/>
    </reference>
    <reference anchor="RFC9457" target="https://www.rfc-editor.org/info/rfc9457">
      <front><title>Problem Details for HTTP APIs</title>
      <author initials="M." surname="Nottingham" fullname="M. Nottingham"/>
      <author initials="E." surname="Wilde" fullname="E. Wilde"/>
      <author initials="S." surname="Dalal" fullname="S. Dalal"/>
      <date year="2023" month="July"/></front>
      <seriesInfo name="RFC" value="9457"/><seriesInfo name="DOI" value="10.17487/RFC9457"/>
    </reference>
    <reference anchor="TXN-TOKENS" target="https://datatracker.ietf.org/doc/draft-ietf-oauth-transaction-tokens/">
      <front><title>Transaction Tokens</title>
      <author initials="A." surname="Tulshibagwale" fullname="Atul Tulshibagwale"/>
      <author initials="G." surname="Fletcher" fullname="George Fletcher"/>
      <author initials="P." surname="Kasselman" fullname="Pieter Kasselman"/>
      <date/></front>
      <seriesInfo name="Internet-Draft" value="draft-ietf-oauth-transaction-tokens"/>
      <refcontent>Work in Progress</refcontent>
    </reference>
    <reference anchor="NIST-SP-800-207" target="https://csrc.nist.gov/pubs/sp/800/207/final">
      <front><title>Zero Trust Architecture</title>
      <author><organization>NIST</organization></author>
      <date year="2020" month="August"/></front>
      <seriesInfo name="NIST SP" value="800-207"/>
    </reference>
    <reference anchor="NIST-SP-800-162" target="https://csrc.nist.gov/pubs/sp/800/162/upd2/final">
      <front><title>Guide to Attribute Based Access Control (ABAC) Definition and Considerations</title>
      <author><organization>NIST</organization></author>
      <date year="2014" month="January"/></front>
      <seriesInfo name="NIST SP" value="800-162"/>
    </reference>
    <reference anchor="NIST-SP-800-82" target="https://csrc.nist.gov/pubs/sp/800/82/r3/final">
      <front><title>Guide to Operational Technology (OT) Security</title>
      <author><organization>NIST</organization></author>
      <date year="2023" month="September"/></front>
      <seriesInfo name="NIST SP" value="800-82r3"/>
    </reference>
    <reference anchor="NIST-AI-100-2" target="https://csrc.nist.gov/pubs/ai/100/2/e2025/final">
      <front><title>Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations</title>
      <author><organization>NIST</organization></author>
      <date year="2025"/></front>
      <seriesInfo name="NIST AI" value="100-2e2025"/>
    </reference>
    <reference anchor="OWASP-AGENTIC" target="https://genai.owasp.org/2025/12/09/owasp-genai-security-project-releases-top-10-risks-and-mitigations-for-agentic-ai-security/">
      <front><title>Top 10 Risks and Mitigations for Agentic AI Security</title>
      <author><organization>OWASP GenAI Security Project</organization></author>
      <date year="2025" month="December"/></front>
    </reference>
    <reference anchor="OWASP-CHEATSHEET" target="https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html">
      <front><title>AI Agent Security Cheat Sheet</title>
      <author><organization>OWASP</organization></author>
      <date year="2026"/></front>
    </reference>
    <reference anchor="CISA-AGENTIC" target="https://content.govdelivery.com/accounts/USDHSCISA/bulletins/41544ff">
      <front><title>Careful Adoption of Agentic AI Services</title>
      <author><organization>CISA and international partners</organization></author>
      <date year="2026" month="May"/></front>
    </reference>
    <reference anchor="ANTHROPIC-ESPIONAGE" target="https://www.anthropic.com/news/disrupting-AI-espionage">
      <front><title>Disrupting the first reported AI-orchestrated cyber espionage campaign</title>
      <author><organization>Anthropic</organization></author>
      <date year="2025" month="November"/></front>
    </reference>
    <reference anchor="MITRE-T1698" target="https://attack.mitre.org/techniques/T1698/">
      <front><title>Indirect Prompt Injection (Technique T1698)</title>
      <author><organization>MITRE ATT&amp;CK</organization></author>
      <date year="2026"/></front>
    </reference>
    <reference anchor="MCP" target="https://modelcontextprotocol.io/specification">
      <front><title>Model Context Protocol Specification</title>
      <author><organization>Model Context Protocol project</organization></author>
      <date year="2025"/></front>
    </reference>
    <reference anchor="DAS-CBEA" target="https://datatracker.ietf.org/doc/html/draft-das-consequence-bound-execution-authority-00">
      <front><title>Consequence-Bound Execution Authority for Autonomous Systems</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026" month="November"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-consequence-bound-execution-authority-00"/>
    </reference>
    <reference anchor="DAS-IEV" target="https://datatracker.ietf.org/doc/draft-das-receipt-gated-iev/">
      <front><title>Receipt-Gated IEV</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026" month="October"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-receipt-gated-iev"/>
    </reference>
    <reference anchor="DAS-HANDLE" target="https://datatracker.ietf.org/doc/draft-das-execution-handle/">
      <front><title>Possession Is Not Authority: Execution Handles, Sink Verification, Atomic Consume and Finality Receipts</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-execution-handle"/>
    </reference>
    <reference anchor="DAS-COMPOSITE" target="https://datatracker.ietf.org/doc/draft-das-composite-execution-finality/">
      <front><title>Partial Commit Is Not Finality: Composite Execution Finality</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-composite-execution-finality"/>
    </reference>
    <reference anchor="DAS-TOOL" target="https://datatracker.ietf.org/doc/draft-das-agentic-tool-binding/">
      <front><title>tool_use Is Not invoke(): Binding Execution-Finality to Agent Tool Calls</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-agentic-tool-binding"/>
    </reference>
    <reference anchor="DAS-REVOCATION" target="https://datatracker.ietf.org/doc/draft-das-finality-bound-revocation/">
      <front><title>Revoked but Still Executable: Closing the Authorization-to-Effect Gap with Finality-Bound Revocation</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026" month="September"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-finality-bound-revocation"/>
    </reference>
    <reference anchor="DAS-PATH" target="https://datatracker.ietf.org/doc/draft-das-consequence-path-completeness/">
      <front><title>When the Gate Can Be Bypassed: Consequence-Path Completeness for Execution Finality</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026" month="September"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-consequence-path-completeness"/>
    </reference>
    <reference anchor="DAS-BOUNDARY" target="https://datatracker.ietf.org/doc/draft-das-agentic-effectuation-boundary/">
      <front><title>When AI Agents Hold the Keys: Threat Model and Execution-Finality Requirements for Autonomous High-Consequence Systems</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026" month="September"/></front>
      <seriesInfo name="Internet-Draft" value="draft-das-agentic-effectuation-boundary"/>
    </reference>
    <reference anchor="DAS-ZENODO" target="https://zenodo.org/records/22082995">
      <front><title>The Internet Solved Communication. It Never Solved Authority</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026"/></front>
      <seriesInfo name="DOI" value="10.5281/zenodo.22082995"/>
    </reference>
    <reference anchor="DAS-REFIMPL" target="https://github.com/sangmdas/Execution-Finality-for-AI-Agents-GPUs-Confidential-Computing-and-Zero-Trust-Automation">
      <front><title>Execution-Finality Reference Implementation: Runnable implementation for AI agents, GPUs, confidential computing, and zero-trust automation</title>
      <author initials="S." surname="Das" fullname="Sangam Das"/>
      <date year="2026"/></front>
    </reference>
  </references>
</references>

<section anchor="faq" numbered="true">
  <name>Reviewer FAQ</name>
  <t>The answers below reflect the author's current understanding, which
  may contain inaccuracies. Critical technical review and correction are
  invited.</t>

  <section numbered="true"><name>Is this not simply a Zero Trust Policy Enforcement Point?</name>
  <t>It is a PEP placed at a particular point and given particular inputs.
  A PEP in <xref target="NIST-SP-800-207"/> enforces access decisions for a
  subject and resource. The profile here additionally requires the
  enforcement point to own the effect credential, recompute the act from
  the final bytes, consume single-use authority with the commit, and gate
  later phases on destination receipts. Where an existing PEP already does
  all of this, the profile adds interoperability, not new protection.</t></section>

  <section numbered="true"><name>OAuth RAR already describes the act. What is missing?</name>
  <t>RAR <xref target="RFC9396"/> carries a structured description of the
  intended act to the authorization server and into the token. It does not
  by itself require the resource server to recompute the act from what it
  will actually commit, to consume authority atomically with the commit,
  or to refuse continuation without a destination receipt. RAR is a good
  carrier for SAE constraints (<xref target="ietf-relation"/>).</t></section>

  <section numbered="true"><name>Is this two-phase commit or a saga under another name?</name>
  <t>Two-phase commit coordinates atomicity among participants that agree
  to commit; it does not decide whether a participant is authorized to
  commit a specific act. A saga coordinates compensations after effects.
  This profile decides authorization at the commit point and treats
  compensation as a new authorized act. Multi-sink atomicity is addressed
  in <xref target="DAS-COMPOSITE"/>.</t></section>

  <section numbered="true"><name>Does this stop prompt injection?</name>
  <t>No. It assumes injection may succeed in changing the agent's intent.
  It limits what a successfully injected agent can make real to effects
  inside the SAE on covered routes. Harmful acts within the SAE remain
  possible.</t></section>

  <section numbered="true"><name>What if the agent holds another credential?</name>
  <t>Then the route using that credential is uncovered. The profile
  requires removing effect credentials from agent runtimes and declaring
  any remaining route as uncovered. This is the largest practical
  dependency (<xref target="DAS-PATH"/>).</t></section>

  <section numbered="true"><name>How can the sink reconstruct the act for complex resources?</name>
  <t>Only the sink that first makes the effect real needs to reconstruct,
  and it already has to construct the effect to perform it. Reconstruction
  is hardest where effects are transitive (triggers, controller
  reconciliation). Where a field cannot be resolved, it is UNKNOWN and the
  act is held. Completeness of reconstruction is an open problem
  (<xref target="limits"/>) and is measured, not assumed.</t></section>

  <section numbered="true"><name>What if the sink or validator is compromised?</name>
  <t>If both are compromised, the controls may fail (A1). Attestation,
  separate administration and multi-party validation reduce the
  likelihood of joint compromise (<xref target="sink-trust"/>); they do not
  eliminate it.</t></section>

  <section numbered="true"><name>How much latency does this add?</name>
  <t>No figure is claimed. Validation is placed at rare consequential
  commits, with policy precompiled so the hot path is deterministic
  verification and a per-handle compare-and-swap. Tail latency must be
  measured against a comparable baseline (<xref target="latency"/>).
  Hard real-time loops are excluded.</t></section>

  <section numbered="true"><name>Is fail-closed always correct?</name>
  <t>No. In physical systems, blocking can itself be unsafe. The safe state
  on validator or store unavailability is defined by the operational safety
  case, and the profile does not override interlocks
  (<xref target="scenario-c"/>).</t></section>

  <section numbered="true"><name>Why trust a destination receipt more than the agent's report?</name>
  <t>The agent, its proxy or its tool may be compromised; the destination
  system of record is a different trust domain that observed the effect.
  A receipt is only as trustworthy as its origin, which is why receipt
  origin is recorded and diversification is recommended.</t></section>

  <section numbered="true"><name>What happens on timeout after submission?</name>
  <t>The act becomes INDETERMINATE, and no retry of the same intent is
  permitted until reconciliation with the system of record using the
  idempotency key establishes whether the effect occurred
  (<xref target="consume"/>).</t></section>

  <section numbered="true"><name>Can an attacker split one large act into many small ones?</name>
  <t>Cumulative ceilings in protected counters apply across acts and across
  delegated agents under the same SAE. Decomposition across different
  principals with different SAEs remains an open problem.</t></section>

  <section numbered="true"><name>How does this work with multiple cooperating agents?</name>
  <t>Delegated agents receive derived SAEs that are strict subsets with a
  shared ceiling, computed by the validator (<xref target="delegation"/>).
  No agent passes its credentials to another.</t></section>

  <section numbered="true"><name>Where does it sit in MCP or similar tool protocols?</name>
  <t>At the tool server that owns the effect credential, for tools whose
  effect is consequential (<xref target="protocol-placement"/>). Read-only,
  non-sensitive tools need not participate. Provider-hosted tools that
  cannot be instrumented are uncovered unless the destination enforces.</t></section>

  <section numbered="true"><name>Does it require a TEE or HSM?</name>
  <t>No. Keys and consume state must be isolated from the agent runtime;
  HSM or TEE isolation is recommended for higher consequence classes. A
  software-only PED provides weaker assurance and should be declared as
  such.</t></section>

  <section numbered="true"><name>Does it address data leaking through the model's own answer?</name>
  <t>Only partially, through the output-channel profile
  (<xref target="output-channel"/>), which limits release by source
  classification and recipient. It does not detect paraphrase of
  protected content whose provenance was not recorded.</t></section>

  <section numbered="true"><name>Why bring this to the IETF?</name>
  <t>Because the act crosses components from different parties, a shared
  representation of the authorized act, of commit-time binding, and of
  destination evidence is a protocol interoperability question. The
  mapping to existing working groups is in <xref target="ietf-wg"/>.</t></section>

  <section numbered="true"><name>How would one show this profile does not help?</name>
  <t>Run the evaluation of <xref target="evaluation"/>. If a strong
  conventional baseline (arm B) blocks the same attack corpus with equal
  coverage and comparable latency, the profile offers no security gain for
  that deployment, and that result should be published.</t></section>

  <section numbered="true"><name>Why not simply bind the token to a key, as DPoP does?</name>
  <t>Key binding stops a stolen token from being used by a different
  holder. It does not stop the legitimate holder, including a compromised
  agent, from using the token for a different act than the one approved.
  The handle binds the digest of the exact act and the sink, so an
  authorized key holder obtains nothing for a different act. Key binding
  MAY be layered on top as a token-theft mitigation.</t></section>

  <section numbered="true"><name>If the handle is stolen, can it be used?</name>
  <t>Only for the single act whose recomputed digest it binds, at the sink
  it names, in its phase, before expiry, and once. A thief presenting it
  with a different act is refused. A thief who can cause the same act to
  be submitted gains nothing beyond what was already authorized.
  Lifetime is kept short (minutes), and holder key binding can further
  restrict presentation.</t></section>

  <section numbered="true"><name>What if the validator and the sink canonicalize differently?</name>
  <t>A disagreement means the digests differ and the handle fails
  verification, which is a fail-closed outcome. The risk is the opposite
  case: two different operations that canonicalize to the same bytes. This
  is addressed by using one canonical form (JCS or deterministic CBOR),
  rejecting duplicate keys and ambiguous encodings, treating UNKNOWN
  differently from empty or default, and maintaining shared
  cross-language test vectors, including negative vectors. Canonicalization
  failure is DENY, never allow.</t></section>

  <section numbered="true"><name>Who decides which fields are load-bearing?</name>
  <t>The consequence-class profile for each sink type, reviewed by the
  owner of the resource. A field is load-bearing if changing it changes the
  consequence (beneficiary account, destination, data scope, target set,
  artifact digest). Ordinary trace identifiers are excluded so equivalent
  operations produce equal digests. An omitted load-bearing field is the
  main source of reconstruction error, so profiles should be tested by
  mutating every field, and the evaluation plan counts any undetected
  mutation as a failure.</t></section>

  <section numbered="true"><name>What does the human approver see, and is it what executes?</name>
  <t>Approval is bound to the CAD digest. The approval interface renders
  the reconstructed act (resolved beneficiary, destination, data scope,
  configuration delta), not the agent's description of it. If the act
  changes after approval, the digest changes and the approval no longer
  applies. This addresses displayed-versus-executed divergence but not
  approver fatigue or an approver who approves a harmful act.</t></section>

  <section numbered="true"><name>Does this replace human oversight or risk management?</name>
  <t>No. It is a technical enforcement mechanism that can implement a
  decision to require approval or to limit scope. It does not decide
  policy, does not evaluate whether an authorized action is wise, and does
  not substitute for organizational risk management or safety cases.</t></section>

  <section numbered="true"><name>How does this relate to the EU AI Act?</name>
  <t>The profile does not make a system compliant and this document makes
  no legal conclusion. For high-risk AI systems, the Act sets
  requirements on risk management, logging and record-keeping, human
  oversight, and accuracy, robustness and cybersecurity. Technical
  measures of the kind described here (enforced scope limits at the effect
  point, authenticated records of what was actually committed, approvals
  bound to the exact act, and defined failure behaviour) may be among the
  measures a provider selects. Whether they are sufficient is a
  determination for the provider against the legal text, harmonised
  standards and guidance, not for this document.</t></section>

  <section numbered="true"><name>Do receipts and logs create privacy or data-protection concerns?</name>
  <t>They can, because they record beneficiaries, recipients and data
  scopes. Objects that leave the protected domain should carry digests
  rather than values, retention should follow the consequence class, and
  receipt logs should not become a general activity record about
  individuals. Data-protection review of retention and access is a
  deployment task.</t></section>

  <section numbered="true"><name>What if the agent regenerates slightly different arguments on retry?</name>
  <t>A different argument set is a different act with a different CAD
  digest, so the earlier handle does not apply. To prevent
  resubmit-until-accepted behaviour, the profile allows only one live
  authority per intent_id. A new attempt for an intent with an AUTHORIZED,
  HELD or INDETERMINATE act is refused until that act is resolved.</t></section>

  <section numbered="true"><name>What is the smallest useful first deployment?</name>
  <t>One high-value boundary with a narrow commit point, such as an export
  service or a payment gateway in a sandbox. Remove the effect credential
  from the agent runtime, reconstruct the act at that boundary in shadow
  mode first, then enforce. Measure attempted versus completed
  unauthorized effects, tail latency and false denials against a strong
  conventional baseline before expanding.</t></section>
</section>

<section anchor="companions" numbered="true">
  <name>Companion Documents and Resources</name>
  <ul>
    <li><xref target="DAS-CBEA"/> -- general consequence-bound execution
    authority.</li>
    <li><xref target="DAS-IEV"/> -- staged effects, evidence-carrying
    receipts and continuation authority.</li>
    <li><xref target="DAS-HANDLE"/> -- Execution Handle, sink verification,
    atomic consume and Finality Receipt objects reused here.</li>
    <li><xref target="DAS-COMPOSITE"/> -- multi-sink, all-or-none
    finality.</li>
    <li><xref target="DAS-TOOL"/> -- binding to agent tool-call
    interfaces.</li>
    <li><xref target="DAS-REVOCATION"/> and <xref target="DAS-PATH"/> --
    revocation at the finality boundary and route completeness.</li>
    <li><xref target="DAS-BOUNDARY"/> -- threat model for autonomous
    high-consequence agents.</li>
    <li><xref target="DAS-ZENODO"/> -- general technical background.</li>
    <li><xref target="DAS-REFIMPL"/> -- runnable reference implementation.
    Results there are environment-specific and are not claims of this
    document.</li>
  </ul>
</section>

<section anchor="prep" numbered="false">
  <name>Preparation Note</name>
  <t>The architecture is the author's. Parts of the analysis, scenario
  descriptions and editorial preparation were AI-assisted. Industry
  references describe published material only and imply no affiliation,
  endorsement, adoption or implementation by any organization.</t>
</section>
</back>
</rfc>
