Consented and Attributable Agent Authority for Operational-Technology Control Actions

Introduction Two bodies of standards work are moving quickly in parallel, and they do not meet. One is agent identity for the enterprise cloud. A software agent that acts for a person or an organisation is being given a verifiable identity and a way to authenticate itself, composing existing web and workload-identity primitives. The web-bot-auth effort specifies how an automated agent authenticates itself over HTTP using HTTP Message Signatures , and it deliberately declines to bind that key to a human principal. This work is real and useful, and it is scoped to general information systems. It does not address operational technology. The other is operational-technology security. Frameworks such as , , and the reliability standards govern the industrial control systems that run the electric grid, water, pipelines, and manufacturing. They require that actors be identified (the identification and authentication control family), that use be controlled (the use-control family), and that consequential actions be auditable. They state these as requirements. They do not specify a wire mechanism by which an agent-originated command carries the proof that satisfies them, and the installed base of control protocols (Modbus, DNP3, and their peers) authenticates a command largely by its position on the network rather than by anything the sender proved. The gap between the two is specific and, at present, unserved: there is no interoperable way for a command issued to a control system on the authority of an agent to carry a revocable, auditable, principal-bound statement of the authority under which it is issued, such that an enforcement point can refuse the command when that statement is absent or invalid. An agent that can write a setpoint to a turbine, open a breaker, or change a treatment dose is a workload whose authority to do so must be provable, scoped, revocable, and attributable after the fact, at stakes where a wrong action is a physical event rather than a corrupted record. This memo specifies that binding. It introduces no new identity mechanism. It composes primitives specified in separate memos into one envelope, the Command Authority Envelope (CAE), that accompanies an agent-originated OT control action, and it specifies the fail-closed behaviour of an enforcement point that evaluates it. Applicability. The operating condition this profile is written for is a plant that must keep its critical services running through a sustained loss of external connectivity, whether that loss is permanent by design or produced by an isolation event. This is the assumed case rather than an exception the profile tolerates. Nothing in the envelope requires a network path beyond the conduit at the moment of evaluation: an encoding MUST be verifiable offline against cached trust anchors with declared staleness bounds (Section 7), which covers the agent key material, the approver directory and the transparency-log checkpoint alike (Section 8). Where a bound is exceeded the enforcement point fails closed on authority, and never on safety (Section 6). Scope. The actions this profile is written for are those requiring an auditable point in time that ties the user, the command and the authorisation together: an operator-requested process stop outside the independent safety path, a setpoint pushed outside normal operating parameters, the starting or stopping of a process. More generally, it addresses deployments where traditional control protocols are in use but additional controls on the authorisation of commands are required. It is not a general mechanism for machine-to-machine communication within a plant operating inside its set boundaries.

A neighbouring proposal, and where this profile parts from it One other Internet-Draft addresses agent authority for industrial control directly. applies an agent-trust transport to Modbus/TCP, OPC UA, MQTT, and CoAP, mandates ECDSA message signing over those protocols, and maps agent trust levels to the Security Levels of . It supplies, in concrete wire form, much of the transport binding this memo defers (Section 7), and a deployment that wants a worked control-protocol encoding today will find one there. On one point the two proposals disagree, and the disagreement is this memo's central claim. gates safety-critical commands on the agent's trust level: a command to a safety-classified point is rejected when the issuing agent presents an insufficient trust level. This memo forbids exactly that (Section 6). A safety function's right to bring or hold the process in a safe state, and its right to refuse an unsafe command on its own criteria, MUST NOT be made to depend on the resolution, verification, or trust level of any agent credential. Gating a safety command on an identity check makes the safety function unavailable precisely when the identity infrastructure is degraded, and that is a safety regression introduced in the name of security. The two proposals can compose on everything below the safety boundary: agent signing, trust-level-to-Security-Level mapping, and the per-protocol envelopes are complementary to the bindings this memo defines. They cannot compose across the safety boundary, and this memo places that boundary where an OT safety case requires it and does not. The unclaimed ground this memo occupies is the coupling: a consented, resolvable, human-principal authority, plus a named-human authorization at the moment of consequence, plus an attributable append-only record, drawn into a single enforcement-point-evaluated, risk-class-graded, fail-closed refusal profile whose defining axiom is that a safety function is never gated on any of it. Each of the five primitives is specified elsewhere. The refusal profile and the safety carve-out are the contribution.

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

Agent:: A software actor that issues a control action to an OT system. An agent is a workload with a discoverable identity, not a human.
Principal:: The human, or the organisation acting through a human, on whose authority the agent issues an action. The principal is the party that issues the authorisation grant (Section 3.3) under which the agent acts.
Approver:: A named, accountable human who signs off on a specific control action at the moment of consequence, per . The approver need not be the principal, and for a consequential action SHOULD NOT be the agent that initiated the action.
Control action:: A request that changes, or commands the change of, the state of a physical process or of a device that governs one: a setpoint write, a breaker operation, a mode change, a dose change. A read-only observation is not a control action for the purposes of this memo. Where a deployment elects to apply this profile to reads, it does so under the Observe class (Section 5).
Conduit:: In the sense of , the communication path between zones across which a control action travels. This profile is enforced at the conduit, by the evaluating function this memo calls the enforcement point.
Enforcement point:: The function, resident on or at the boundary of a conduit, that evaluates the CAE of an agent-originated control action and refuses the action on any missing or invalid binding. This memo gives the function its own name because uses "conduit" for a channel grouping rather than for an evaluating function, and the two need to be distinguishable in a sentence. The enforcement point is the conduit's evaluating function; it is not a different place.
Command Authority Envelope (CAE):: The structure defined in this memo that a control action MUST carry to be accepted by an enforcement point that implements this profile.
Authorisation grant:: A scoped, revocable object, signed by the principal, that authorises a named agent to perform a named control verb on a named asset until a stated expiry (Section 3.3).
Binding moment:: A named human's signed authorization of one exact control action at the moment of consequence, carried as human-authorization evidence (Section 3.4). The evidence is the receipt; the interaction that produces it MAY be delivered through the briefing-and-binding envelope of .
Risk class:: The category assigned to a control action by its potential physical consequence, which determines which bindings the CAE MUST carry.
Safety function:: A function whose purpose is to bring or hold the process in a safe state, including a safety-instrumented system (SIS). Safety functions are explicitly outside the authority path of this profile (Section 6).

The Command Authority Envelope A control action issued on the authority of an agent to an enforcement point that implements this profile MUST carry a Command Authority Envelope. The CAE is a signed structure carried alongside the control action. The bindings it carries are listed below; the mapping from each binding to the composed artefact that supplies its concrete fields is in Section 7. The CAE binds five things.

Agent identity The CAE MUST identify the issuing agent by a discoverable identifier whose key material is resolvable and verifiable independently of the enforcement point. A deployment reachable from public DNS SHOULD resolve the agent identifier per , for which verification is DNSSEC-rooted and fails closed when DNSSEC is absent. The agent's request signature MUST be verifiable per , consistent with . This binding answers "which machine issued this", and nothing more; on its own it is insufficient, which is the gap the web-bot-auth effort leaves open by design.

Principal reference The CAE MUST carry a reference to the principal on whose authority the agent acts. The reference is a resolvable identity handle, not a bare string. This binding is the one the agent-authentication layer deliberately omits: it names the human behind the machine. A control action whose CAE names no principal MUST be treated as principal-less and refused at any risk class above the lowest (Section 5).

Authorisation grant The CAE MUST carry a reference to a scoped, revocable authorisation grant, issued and signed by the principal, that authorises this action. The grant MUST name the specific asset (the zone, conduit, device, or point), the specific control verb it authorises, and the specific agent authorised to exercise it, and it MUST carry an expiry. A grant that names a broader scope than the action does not satisfy this requirement more strongly; it satisfies it exactly to the overlap, and an enforcement point MUST evaluate coverage against the specific action, not against the grant's breadth. Authority is captured against the action, not inferred from an operator's one-time enrolment. This section states conformance requirements, not one mandated object. Any grant artefact meeting the requirements below is a conforming filler, and this memo names the fillers known to it without preferring one. A conforming authorisation grant: G1. is signed by the principal whose authority the action invokes; G2. names the asset, the control verb, and the authorised agent; G3. carries an expiry; G4. is content-addressed, so that a record can reference exactly the grant that was in force; G5. is revocable, and permits an enforcement point to establish its revocation state or to refuse for want of it; and G6. permits an enforcement point to evaluate coverage against the specific action and to fail closed, under a distinguishable reason, on any requirement it cannot meet. Two fillers are known at the time of writing, and neither is privileged by this memo. EP-CONSENT-GRANT-v1 meets G1 and G3 through G6, and needs an authorised-agent profile and check before it meets G2. It is a signed standing grant naming an asset, a control verb, and an expiry, content-addressed by a grant hash, revocable by a revocation statement against that hash, and its verifier refuses under distinct reasons for signature, validity window, revocation, asset, verb, and grant-binding failures. G2 requires the grant to name the authorised agent, and this filler does not yet define or check that binding; its asset, verb and validity checks are unaffected. Current revocation status at an enforcement point depends on that point's freshness policy (Section 8), and offline signature verification alone does not establish it. maps it against the authorisation-grant, binding-moment and audit-record bindings of this memo. It is, so far as the authors are aware, the first artefact specified and implemented against this row. The grant of does not meet G2 as it stands. It authorises a reader to READ an attribute about a data subject and is signed by that subject, so it grants a read of an attribute rather than a command over an asset, and it requires profiling with asset, verb, and authorised-agent fields before it fills this row. The distinction is worth stating plainly, because the two objects are easily confused. The grant of this section authorises an agent to ACT on an asset and is signed by the principal whose authority the action invokes. The grant of authorises a read and is signed by the data subject. The two carry the same revocation and content-addressing machinery and the same signed-by-the-authorising-party discipline; they differ in what they authorise and in who signs. The principal of this memo and the subject of are different roles occupying the signer slot of two different objects.

Binding moment For a control action whose risk class requires it (Section 5), the CAE MUST carry an authorisation artefact that binds a named, accountable human to this exact action before the action executes. The artefact is bound into the CAE as human-authorization evidence per , and the bound evidence is an authorization receipt per , or an equivalent artefact carrying the same properties. The artefact MUST satisfy, at the enforcement point, the binding requirements of :

it is credited only against artefact bytes, a signature the enforcement point verifies or a digest it matches, and never against a bare assertion that a human approved (B1, digest grounding);
its action binding MUST agree with the control action's asset and verb, so that an approval issued for one action cannot be spent on another (B2, action agreement). An artefact authorising a different action MUST invalidate the binding, not merely weaken it;
the enforcement point MUST distinguish verifying the artefact (its digests and signatures hold) from accepting it (its issuer's key is pinned out of band), and MUST NOT accept an artefact from an unpinned issuer (B3, verified versus accepted);
the absence of the artefact is the absence of evidence, never a default grant of authority (B4, fail-closed absence);
where the artefact is available in both an embedded and a referenced form, the two MUST agree, and an enforcement point that resolves the reference MUST refuse the action if the resolved artefact differs from the embedded one (B5, form consistency). Section 7 offers both forms, so this requirement is live in this profile.

The action the enforcement point evaluates and the action it forwards MUST be the same action. An enforcement point MUST fix the action's asset, verb and parameters before it verifies the artefact, and MUST forward that fixed action to the process. Where verification, revocation lookup or spend recording introduces a wait, the action MUST NOT be re-read from a mutable structure after that wait. An enforcement point that verifies one object and forwards another satisfies B2 against an action nobody authorised. B2 alone does not reach this, because agreement is checked at one moment and the action executes at another, and the interval between them is where the substitution happens. The receipt MUST carry the approving human's own signature over a digest of the action, be verifiable by the enforcement point offline against cached key material and a published log checkpoint per , and MUST NOT be usable more than once. A binding moment MUST carry an expiry. An enforcement point MUST refuse an artefact presented after its expiry, and MUST NOT treat a captured approval as valid for any window the artefact does not itself state. A captured decision is evidence of a decision taken at a moment, not a standing token.

Audit record The CAE MUST carry an append-only, independently attributable record of the action. The record is a signed statement per the SCITT architecture , registered on a transparency service, with a COSE receipt proving the statement's inclusion in the service's append-only log. The statement MUST be sufficient to attribute the action afterward: which agent, on which principal's authority, under which authorisation grant, with which binding moment where one was required, against which asset, at which time per . Where a binding moment was required, the statement MUST reference the authorisation artefact by digest per , so that the record of the action and the evidence of its human authorisation bind the same bytes. This record composes published standards. SCITT signed statements and COSE receipts are the append-only, independently verifiable artefacts the evidence requirements of and call for. The agent-authentication layer references such a record but does not itself produce it; that gap, not the absence of any suitable format, is what this binding fills. The record MUST carry a provenance term drawn from the closed vocabulary of , stating how the assertion the record makes was corroborated. The term labels the record, which is an assertion about the action; it does not label the physical action itself, which is attributed by the signed statement. This distinction matters, because a closed provenance vocabulary can say how a claim came to be believed and cannot say that a valve moved. Where the agent includes a machine-generated rationale in the record, for example a stated reason for escalating the action to a human in the sense of the initiator attestation of , that rationale carries its own term from the same vocabulary. An enforcement point MUST refuse a control action whose audit record carries no provenance term, and MUST refuse one whose term has decayed to uncertainty. A record that cannot say how it came to be believed does not confer authority. The record states what the enforcement point observed, not what it intended. Where the outcome of the action is unresolved in the sense of Section 4, the record MUST say so, and MUST NOT assert that the action took effect. A statement that an asset reached a state, written without observing the process, corroborates nothing, and the provenance term of labels the record rather than repairing it.

Enforcement-Point Evaluation and Fail-Closed Behaviour An enforcement point that implements this profile MUST evaluate the CAE of every agent-originated control action before the action reaches the process, and MUST refuse the action if any binding required for the action's risk class is absent, malformed, expired, revoked, or unverifiable. Refusal is the default and the safe state for authority. An enforcement point MUST NOT accept a control action on the ground that the CAE could not be evaluated (for example because a revocation status could not be reached); an unevaluable authority is a refused authority. This is the same posture as the gate: the enforcement point refuses the request rather than attempting to prove, cryptographically, that the agent lacked authority. That is an honest and contestable trust boundary, and Section 8 states it as such. Evaluation before the action yields one of two outcomes, accept or refuse. The action itself yields a third. Where an enforcement point has accepted a CAE, dispatched the action, and cannot determine from the process whether the action took effect, the outcome is unresolved. An enforcement point MUST NOT report an unresolved outcome as performed, and MUST NOT report it as refused. Reporting it as refused is the worse of the two, because a refusal asserts that nothing reached the process. An unresolved outcome carries no authority forward. Retrying the action is a new control action, requiring its own CAE and, where the action's risk class requires a binding moment, its own binding moment (Section 5). An enforcement point MUST NOT re-present the artefact of an unresolved action, and MUST NOT resolve the outcome by asking the agent what happened. Fresh authorization alone MUST NOT permit another attempt at the same unresolved operation. The deployment MUST retain a stable operation identifier and a durable unresolved state across restart and re-authorization. It MUST NOT release another attempt merely because the original response was lost. Authenticated reconciliation or a deployment-specific recovery procedure must establish whether, and under what conditions, another attempt is permissible. Section 3.4 requires that a binding-moment artefact MUST NOT be usable more than once, and the enforcement point is what makes that hold. An enforcement point MUST record each artefact it accepts as spent, and MUST refuse any later action presenting an artefact already recorded. Two records carry that requirement rather than one. The spend record at the enforcement point MUST survive a restart of the enforcement point, and where a downstream conduit can duplicate an action already dispatched, the record that suppresses the duplicate MUST survive a restart of that conduit. Survival of power loss is a requirement on the deployment, and no implementation reported in Section 12 demonstrates it. Spend state held only in volatile memory makes the requirement once per uptime rather than once, and a power cycle is an ordinary event in an OT deployment rather than an exceptional one. An artefact dispatched against an unresolved outcome is spent. Single spend holds only where one enforcement point evaluates every agent-originated control action for a given asset. Where more than one enforcement point can admit an action for the same asset, no single enforcement point can hold the spend record that Section 3.4 requires, and a deployment MUST either scope each asset to a single enforcement point or refuse the action, on the same ground as an unreachable revocation status. What this section requires of an enforcement point is at-most-once admission and forwarding of the authorised attempt. It does not require, and cannot deliver, exactly-once physical effect. An unresolved outcome means the process may or may not have acted, and no property of the authority path resolves that question. At-most-once admission and reconciliation alone do not establish exactly-once physical effect. That depends on the executor, the device, and how uncertain outcomes are resolved. A deployment claiming exactly-once effect on the strength of this profile alone has claimed something this profile does not provide. Refusal of a control action on authority grounds MUST NOT itself be able to prevent, delay, or gate a safety function (Section 6). The authority path and the safety path are separate, and the profile lives only in the former.

Risk Classes An enforcement point assigns each control action a risk class by its potential physical consequence. The mapping from action to class is a property of the deployment and its process hazard analysis, not of this memo; this memo specifies only which bindings each class requires. A deployment SHOULD align its classes with the Security Levels of . Three classes are defined; a deployment MAY define finer gradations between them.

Observe (lowest):: A read of process state, carried under this profile only where a deployment has elected to apply the profile to reads (Section 2). Where it applies, the CAE MUST carry agent identity and an audit record. Principal reference, an authorisation grant, and a binding moment are OPTIONAL.
Adjust (middle):: A change within a bounded, pre-authorised safe envelope, for example a setpoint move within an interlocked range. The CAE MUST carry agent identity, principal reference, an authorisation grant covering the asset and verb, and an audit record. A binding moment is RECOMMENDED and MAY be required by the deployment.
State-change (highest):: A change of process or device state with safety or reliability consequence, for example a breaker operation, a mode change, or a change that leaves an interlocked envelope. The CAE MUST carry all five bindings, and the binding moment MUST be present and valid.

An enforcement point MUST refuse a State-change action whose CAE lacks a valid binding moment, without exception, and MUST NOT downgrade an action's class to avoid a binding requirement. Where a State-change action's outcome is unresolved (Section 4), a retry is a new action of the same class. Its binding moment MUST be present and valid in its own right, which means a human decides again. That decision is necessary where this section requires it and is not sufficient: the retry remains subject to the rule in Section 4 on another attempt at an unresolved operation, and a fresh binding moment does not by itself clear the uncertainty.

Safety Carve-Out This is the requirement the profile refuses to compromise, and it is stated here, ahead of the security considerations, because it is the one an OT engineer will test first. A safety function MUST NOT be gated on any binding in this profile. A safety-instrumented system, an emergency shutdown, a hardware interlock, a protective relay operating on its own criteria: none of these is an agent-originated control action in the sense of this memo, and none of them MAY be made to depend on the resolution, verification, or revocation status of a CAE. A safety action that a plant would take autonomously MUST remain takeable when every network, every DNS resolver, and every consent endpoint is unreachable. This is where this profile and part (Section 1.1). A design that rejects a safety-classified command because the issuing agent presented an insufficient trust level has placed an identity check in the safety path. This memo forbids that placement. The profile constrains who may command a process to move. It has no authority over the process's own right to protect itself. A design that allowed an identity check to block a trip would be a safety regression introduced in the name of security, and this memo forbids it.

Encoding and Transport Binding This section maps each CAE binding to the composed artefact that supplies its concrete fields. It stops there deliberately. This memo does not mandate, and this revision does not specify, a single outer CAE encoding or a novel wire structure of its own. Where the profile needs a field, it takes it from a primitive already specified elsewhere; where a concrete encoding decision remains, it is named as such and left to a later revision. The CAE is a signed structure. Each binding is filled as follows.

CAE binding	Filled by	Concrete fields come from
Agent identity (3.1)	Resolvable agent identifier and request signature	for the identifier and its DNSSEC-rooted key material; for the signature, consistent with
Principal reference (3.2)	Resolvable principal handle	A resolvable identity handle naming the human on whose authority the agent acts
Authorisation grant (3.3)	Principal-signed, scoped, revocable grant naming asset, verb, agent, and expiry	Any filler meeting G1 to G6 of Section 3.3. meets G1 and G3 to G6 and needs an authorised-agent profile and check for G2; the grant structure of requires profiling with asset, verb, and authorised-agent fields first
Binding moment (3.4)	Named-human authorization evidence carrying an approver's signature over the action digest	The binding object (`human_authorization_ref` by digest, or `human_authorization` embedded) of , carrying an authorization receipt per ; the interaction optionally via
Audit record (3.5)	SCITT signed statement with a COSE inclusion receipt	A signed statement per , registered on a transparency service, with a receipt per ; the binding moment referenced by digest per

An encoding of the CAE MUST meet the following requirements, which are properties the OT environment imposes and are independent of the field map above. An encoding MUST be verifiable offline against cached trust anchors, because many OT environments are segmented from public networks for long, declared intervals (Section 8). The offline-verifiable authorization receipt of and the offline COSE inclusion proof of are chosen for this reason. An encoding MUST carry a freshness element (a nonce and an timestamp with a declared maximum age) to bound replay. An encoding SHOULD ride above, and MUST NOT weaken, the transport security of the underlying session; where the session is , the CAE rides above the OPC-UA secure channel, which proves the channel while the CAE proves the authority. Transport bindings for specific control protocols are out of scope for this revision. specifies per-protocol signed envelopes for Modbus/TCP, OPC UA, MQTT, and CoAP, and a deployment MAY carry a CAE within such an envelope; the two are complementary below the safety boundary (Section 1.1). The concrete outer encoding of the CAE, how the five bindings above are serialised together into one structure, is the natural content of a companion document or a future revision.

Security Considerations This section is written to be attacked. Several of the boundaries below are honest and contestable rather than closed, and they are marked as such. Independent review from an operational-technology and critical- infrastructure background is the review this document most needs. Availability over authentication. In OT the priority order is availability, then integrity, then confidentiality, the inverse of the usual information-systems order. This profile is built to that order: it fails closed on authority and never on safety (Section 6), and it refuses rather than blocks. The reviewer should test whether any path in a deployment could let an authority check stall a time-critical control loop; if one exists, the deployment has mis-placed the gate. Command integrity and diverse-channel confirmation. The CAE binds authority to a control action and makes the action attributable; it is not, on its own, an integrity mechanism for the command value on the wire. The profile requires that an encoding ride above and not weaken the transport security of the underlying session (Section 7), so the command value is protected to the integrity the session provides, and no further. Where a corrupted or spoofed command value is itself a hazard, and in particular where the controlled function carries a safety-integrity requirement at SIL 2 or above in the sense of and , the transport integrity of a single command path is not sufficient by itself: a deployment SHOULD confirm the commanded state over a channel independent of the command path, for example an independent read-back of the achieved process state, and SHOULD treat a discrepancy as a fault for the process's own safety logic to handle rather than for this profile to handle. Providing such a diverse channel, and meeting a stated SIL target for the end-to-end control function, is a functional-safety engineering task governed by and ; it is substantial work, it is a property of the deployment and its safety case, and it is out of scope for the authority binding this memo specifies. This memo neither supplies nor weakens that integrity: a diverse-channel confirmation composes beneath the authority profile, and, like every mechanism here, it MUST NOT be placed where it can gate a safety function (Section 6). Refuse, do not prove. An enforcement point refuses an action whose authority it cannot verify. It does not prove the agent lacked authority. This is a deliberate, contestable boundary inherited from . An adversary who can make a valid CAE unevaluable can cause refusal, which in an availability-first setting is itself a denial-of-control concern; the mitigation is the offline-verifiable trust anchor and cached revocation state below, and the reviewer is invited to find the residue. Binding-moment forgery and the human-in-the-loop. The gravest failure this profile must exclude is a State-change that executes on no human decision. The requirement that the binding moment be an authorization receipt carrying the approver's own signature over the action digest, verifiable offline and not usable more than once (Section 3.4), is what excludes it: an agent cannot manufacture that signature, and cannot replay a genuine one. One residual surface remains and is stated plainly: the presentation attack of Section 11.3, in which an approver may sign a faithful-looking rendering of the wrong action. This memo inherits that residual and the mitigations states (render from the hashed bytes, register render templates under the policy, and for the highest classes render the material parameters on a surface the orchestrating operator did not author). It is an enforcement-side obligation this profile places on the enforcement point, not a property it can assume. Revocation latency versus plant time. An authorisation grant revoked mid-session MUST stop future actions it covered within a bounded, declared latency. In a plant, that latency competes with real-time control constraints and with intervals of network segmentation. The trade between revocation freshness and offline operability is real and is not fully closed here; a deployment MUST declare its revocation latency budget and its maximum trust-anchor staleness, and MUST NOT let either gate a safety function. A revocation cannot recall an action already released to the process, and this profile does not close that gap. Key distribution in segmented plants. DNSSEC-rooted discovery per assumes the resolver is reachable. A segmented or air-gapped plant is not. This profile therefore requires offline verification against cached trust anchors with declared staleness bounds, for the agent key material, the approver directory of , and the transparency-log checkpoint of . The management of those anchors, their rotation, and their revocation across a fleet of long-lived devices is the same lifecycle problem that current OT security guidance identifies as largely unsolved, and this memo does not claim to solve it; it requires only that a deployment state its bounds and fail closed on authority when a bound is exceeded. Confused deputy and compromised agent. A valid CAE proves authority, not intent. A compromised agent holding a valid grant can issue any action the grant covers. The mitigations are scope minimality (a grant naming the exact asset, verb, and agent, Section 3.3), the binding moment for consequential classes (Section 3.4), which a compromised agent cannot forge because it carries a human's own signature, and the audit record (Section 3.5) that makes the action attributable after the fact. None of these prevents a first malicious action within scope; they bound its blast radius and guarantee its attribution. Operator as adversary. Consistent with the wider architecture this profile belongs to, the operator of the identity and consent infrastructure is treated as a potential adversary. The authorisation grant is signed by the principal and not the operator; the binding moment is signed by an approver key the operator does not hold (Section 3.4, and the corresponding guarantee in ); the audit record is a SCITT signed statement on an append-only log whose checkpoint the operator cannot rewrite undetectably (, ). These exist so that no single operator is structurally required and every action is visible and attributable, rather than trusting the operator to behave. Scope and the deliberate omission. This memo specifies only the binding and refusal semantics over already-specified discovery, authorisation, human-authorization, and transparency primitives. The methods by which a principal's identity or trustworthiness is inferred are out of scope by construction, and no such method is described, referenced in detail, or required here. A reviewer does not need those methods to judge the trust model, the fail-closed behaviour, or the safety carve-out, which are the parts that matter for this document.

Deployment and Incremental Adoption This profile is meant to be adopted incrementally, alongside the installed base rather than in place of it. The enforcement point (Section 2) is an added function at a conduit boundary; it does not replace a control protocol, a safety system, or a historian, and a deployment can introduce it for one asset class or one high-consequence verb before extending it. The authority bindings compose above existing per-protocol transports, including the signed control-protocol envelopes of where those are deployed, so a site that has already invested in a transport binding keeps that investment. Adoption of a profile like this depends on the integrating parties, the control-system vendors, the asset owners, and the operators of the discovery, authorisation, and transparency primitives, each investing in the integration on its own side, and a specification alone does not create that investment. This memo takes the position that the incentive most likely to carry that investment without a mandate is that the profile discharges an obligation the deployer already holds rather than adding a new one. and the reliability standards already require that consequential actions be identified, use-controlled, and auditable; a deployer meets those requirements today with bespoke, non-interoperable, and often manual evidence. The append-only, independently verifiable audit record this profile carries (Section 3.5) turns that standing, unfunded compliance obligation into a concrete and reusable mechanism, which lowers an existing cost rather than imposing a new one. The concrete integration path on each party's side, and the commercial and operational incentives that make a given party invest, are deployment matters beyond the scope of this memo and are the natural content of a companion deployment document.

IANA Considerations This document has no IANA actions in this revision. A future revision that specifies a concrete CAE encoding is expected to register a media type and MAY request registries for binding types and risk-class identifiers, per .

Normative and Informative References Several references in this document are normative because an implementer requires them to construct or verify a Command Authority Envelope, yet they are individual Internet-Drafts rather than published standards: the Morrison-family memos , , and , and the Schrock EMILIA-Protocol memos and . A normative reference to a work in progress will hold this document in the RFC Editor queue until the referenced drafts are published or the references are re-scoped; the authors acknowledge this and expect to revisit the normative and informative split as the referenced work matures. The remaining normative references are to published standards: , and the transparency pair (SCITT) and (COSE Receipts).

Implementation Status This section records the status of known implementations in accordance with . It is intended to assist the IETF in its decision processes for this document. The description of implementations in this section is intended neither to describe those implementations as complete or correct nor to endorse them; the listing of an implementation here does not imply endorsement by the IETF. This section is expected to be removed before the document advances beyond the Independent Stream. No independent implementation of CAE evaluation at an enforcement point is known at the time of this revision. An independent implementation, against one concrete control-protocol binding, remains the strongest near-term signal this document could receive, and is explicitly solicited. Several of the composed primitives do have running code, and two results are reported here. The EMILIA Protocol implementation (, ) reports three same-team reference ports, in JavaScript, Python and Go, agreeing across 21 suites and 332 conformance vectors at revision 8f0d9a9fec507a70ea006a91c0765e539a96ffaa, together with an externally authored Rust implementation built from a pinned public source tree that matched 16 suites and 164 of those vectors at that pin; that Rust result is recorded at revision 2535c192553903c63bdc56499da8db7b5d719afc, which fixes the Rust source at 7faba36010e7590727bebbc5b9dcceee60539b9b and the evaluator run at 18739076c822bdc757147326e8d36716432d1b41. This Rust result is time-pinned and is not yet a strict clean-room acceptance result. Each revision named in this paragraph was supplied by the implementer and is reported as supplied; none was independently re-run for this document. Separately, the canonicalisation on which digest grounding depends (B1, Section 3.4) is reported by two independently written implementations, one in Python and one in TypeScript, agreeing on a vector suite of 16 canonicalisation cases and 5 required refusals. Nineteen of the twenty-one are exercised on both sides. Two of the refusals run on the Python side only: a parsed JavaScript value cannot present a duplicate member name, and an integer outside the IEEE 754 double range is outside the range RFC 8785 specifies. Both implementations' test suites read that suite from a single file, and the suite was derived once and then re-derived independently on the other implementation before being accepted. The result reported here was obtained on 18 September 2026 against the then-current revision of both implementations, neither of which is public at the time of writing. This is agreement on what the two sides hash and not on CAE evaluation, and it is reported because a digest-grounded binding is only as interoperable as the bytes the two sides hash. Two implementations that disagree about an action's canonical form disagree about every binding computed over it, and do so silently: the signatures verify on each side and match on neither. The deployment claims this section previously carried for , and are withdrawn from it. One of them described an architecture that was subsequently replaced, and the others have not been re-verified against the code that would substantiate them. A claim about another document's implementation belongs in that document, where its own author checks it; repeating it here gave it a second home that nobody was checking.

Contributors The separation held in Sections 3.3 and 3.4, under which a per-action artefact at the binding moment never rounds up into a standing authorisation, was settled in exchange between the authors. Sections 3.3, 3.4, 4 and 12 of this revision carry text originating with I. Schrock, who was credited as a contributor in earlier revisions and is an author of this one.