Executive takeaway
Use ACS for the control interface; assess the operated deployment against Governed Autonomy. ACS supplies agent-to-Guardian hooks, decisions and evidence mechanisms. It does not by itself establish deployment conformance to this doctrine.
R14 is the consequential distinction. ACS defaults to proceeding on two separate failure paths: a failed startup handshake, and a missing usable runtime decision. Setting runtime on_decision_failure: deny does not cover startup. A deployment seeking to meet R14 must withhold ordinary effects when required checks cannot be satisfied, or use pre-authorized bounded continuity controls that meet the applicable baseline. An audit record alone does not make continued operation conformant.
Check both paths, not just the switch. Assess startup refusal and runtime denial separately, along with any declared continuity mode. Do not confuse either failure path with an explicit Guardian DENY or an expired DEFER, whose timeout decision defaults to deny (ACS §4.1, §6.4, §6). This is a document mapping, not an implementation test or certification.
The OWASP Agent Control Standard (ACS) is an open specification from the OWASP GenAI Security Project. An AI agent, which ACS calls the Observed Agent, sends a structured hook request to a separate Guardian Agent before it acts. The Guardian returns a verdict, and the agent is required to honor it. Trace events and a component inventory let the exchange be reconstructed later. In the project’s own words, ACS defines “how agent platforms expose middleware hooks and how open-source tooling enforces safety policy through those hooks.”
This doctrine works at a different layer. The mandatory runtime baseline of Doctrine v4, R01 to R18, states obligations for a deployment within a declared boundary, and the Assessment Companion, R19 to R24, states how a claim about that deployment is examined. The doctrine defines no wire format. ACS defines mandatory interface and conformance obligations through ACS-Core and the profiles a deployment declares; it does not supply the full deployment baseline assessed here. This page maps every baseline requirement to the ACS provisions that bear on it, says for each whether the two share a rule, whether ACS supplies a mechanism toward it, or whether ACS leaves it to the deployment, and records the one place where their defaults point in opposite directions.
What was read. Read on 14 November 2026: the ACS documentation site at owaspacs.org/docs, which presents itself as release 0.1.4 of specification v0.1.0, specifically the Instrument specification, the conformance profiles, the Trace events, the hook catalog and the Inspect (AgBOM) page; and the repository, including its README, CHANGELOG, LICENSING and the conformance claim template. The documentation site tracks the repository’s main branch, which carries normative wording merged after the 0.1.4 tag; this page follows the published text and notes the one difference that matters here. The compared documentation snapshot is repository main commit 5db28a91728ea7678c85fb93a909f92dffe1a540; release tag v0.1.4 resolves to 834f3f0d51e7bbbd9a68ff682bec89f97ddd72d8. These immutable references identify the source versions; the documentation links below remain navigation aids to the published site. Not read: the reference implementations, the schemas themselves, the A2A hook pages, the identity working documents beyond their status notes, and the project’s issue tracker. Nothing on this page implies OWASP endorsement of this doctrine, or the reverse. Where this page and ACS disagree about ACS, ACS is right.
Status and license of the source. Pre-1.0. The CHANGELOG states that pre-1.0 releases “MAY change the ACS-Core baseline and the Trace mappings in patch versions.” Code and machine-readable schemas are licensed Apache-2.0 and prose documentation CC BY-SA 4.0, per the repository’s LICENSING file, which also records that releases up to and including v0.1.0 were published under the MIT License. This page quotes ACS documentation briefly for commentary and does not adapt it.
How to read the mapping
Each row names a requirement of this doctrine, the ACS provisions that bear on it, and one of four relationships:
- Shared rule. ACS states, for its own layer, the same rule the requirement states. Both documents refuse the same shortcut.
- Mechanism. ACS defines a mechanism a deployment can use to produce evidence toward the requirement. The requirement itself, and the policy content that makes the mechanism meet it, stay with the deployment.
- Opposite default. ACS permits, by default, behavior the requirement does not allow without further declaration.
- Not addressed. ACS v0.1 does not cover the requirement. Some of these are deferred to a later ACS version by ACS’s own roadmap.
No row is an equivalence. In the Assessment Companion, “Equivalent control accepted” is a claim state that an assessor other than the deployment’s accountable owner reaches for a named deployment (R23). A mapping between documents cannot reach it, and nothing below should be read as doing so. The doctrine’s conformance criteria sit above any interface a deployment uses to meet them: as the doctrine puts it, “A product’s capabilities do not establish conformance of a system operated with it.”
The mandatory runtime baseline, R01 to R18
| Requirement | ACS provisions | Relationship and what remains with the deployment |
|---|---|---|
| R01 Declare the deployment boundary | ACS-Inspect: an Agent Bill of Materials snapshot per session, serialized to CycloneDX, SPDX or SWID; ACS-Inspect-Dynamic: agbom/changed on mid-session component changes (conformance, Inspect). Both are optional profiles. | Mechanism. The AgBOM can supply the “model, tool and integration bindings on which its claims depend” that R01 asks the profile to name. Owners, missions, effect classes, enforcement points and excluded paths remain the deployment’s declaration. ACS-Core records its own coverage gap in the same spirit as R01’s rule that “An uncovered path shall not be represented as governed”: an MCP server added mid-session under a no-MCP handshake is not wrapped until the next session (ACS-Core, Wrapped MCP). |
| R02 Bind accountable identities | Authenticated handshake and per-message signature; approver authentication is required and the Guardian “MUST verify approver identity against policy” (§9); provenance fields populated “never by the LLM” (§1). | Mechanism. ACS keeps attribution off the model’s output path, which matches R02’s refusal to accept “model narration as attribution.” The grant chain R02 requires (subject, issuer, audience, scope, validity and delegation limits) is not specified: the ACS identity material describes itself as a working document that “does not propose a wire format or normative requirements yet.” |
| R03 Protect mission records | Intent and SessionContext: an append-only, hash-chained context with a published chain head; an approver’s intent extension is recorded with the approver’s identity and is “the only conformant path to mutate Intent.parsed after Intent is committed” (§9.1). | Shared rule within a session: committed scope does not widen silently. Intent is optional in ACS-Core (normative only when intent-based enforcement is the paradigm), and a mission record that persists across sessions, with resource scope and versions, remains the deployment’s. |
| R04 Verify relationships and freshness | DEFER with reasons insufficient_context and pending_dependency (§6). | Not addressed as a requirement. ACS gives a Guardian a way to say a verdict is not yet reachable. Which authoritative sources establish a mission-to-resource relationship, and their freshness rules, are Guardian policy content ACS does not define. |
| R05 Declare policy and fact uncertainty | “Facts on the wire; classification in policy” (§1); DEFER reason low_confidence and an evaluator confidence field (§6.1). | Mechanism, narrow. ACS separates recorded facts from trust classification, consistent with R05’s rule that provenance evidence “shall not be represented as proof of truth.” The profile-level declaration of assumptions, owners and escalation conditions is not addressed. |
| R06 Bound effects outside the agent | Separate Guardian with a deterministic layer that runs first; “The agent MUST NOT have knowledge of hooks” (§1, §2); hooks before tool calls, memory writes and sub-agent starts (hooks). | Mechanism. The hook points are where R06’s external controls can sit, and the separation supports “The acting agent shall not be able to alter its governing policy or controls.” R06 also requires external controls to “deny unknown, unclassified or unenforceable effects.” ACS leaves policy strictness to the deployment: “A permissive Guardian is a conformant but permissive deployment, not a violation” (ACS-Core). |
| R07 Preserve valid authority through execution | Decision honoring: “A hook is a control point only if the Observed Agent waits for the verdict and applies it,” and a framework that proceeds without applying it “is not conformant”; a MODIFY the agent cannot apply is treated as DENY except for postCompact, where the client keeps the original summary and records an audit event (the Guardian substitutes ALLOW) (§6.4, §6.5); ASK routes to an authenticated approver (§9). | Shared rule for binding: R07 requires that protected enforcement “bind the evaluated effect and relevant conditions to what executes,” and ACS makes the same demand of the agent. Validity windows, effect-time rechecks for queued or delayed work, and the rule that prior success does not establish trust remain with the deployment. |
| R08 Constrain delegation across authorities | subagentStart is required for sub-agent-capable clients; without it, ACS says, a Guardian “is blind to cross-agent propagation” (ACS-Core). Recursive ASK, quorum, A2A wrapping and multi-tenant isolation are deferred to v0.2 (§15). | Not addressed in v0.1, by ACS’s own scoping. A Guardian can see a sub-agent being started. What authority the sub-agent carries, and how obligations, budgets and onward-delegation limits pass along the chain, is R08 and Plane 5, and is where ACS plans its next version. |
| R09 Resolve authority conflicts explicitly | DEFER reason conflicting_policies; an expired DEFER takes its timeout_decision, which defaults to deny (§6); a decision may cite several policy_references (§6.1). | Mechanism. ACS gives escalation of an unresolved conflict a wire shape, and its expiry fails closed. The “declared rule” that composes authorities, which R09 requires, is the deployment’s. |
| R10 Define effective revocation | No revocation method in the v0.1 specification. | Not addressed. Effectivity and distribution bounds across queued, in-flight, delegated and irreversible effects remain the deployment’s. |
| R11 Control cumulative effects | Support for cumulative-context denial paradigms through cited_provenance_ids and policy_data (§6.2); cascading deferrals “MUST be bounded per session” (§6). | Mechanism, within one session. Limits across agents, delegations, retries and time windows, with shared counters that prevent handoff-based resets, are not addressed. |
| R12 Track obligations and omissions | A per-session count of decision failures at sessionEnd, which ACS recommends (§6.4). | Not addressed. ACS governs steps an agent attempts. Required outcomes, deadlines and missed obligations, which R12 says governance shall detect “not assess safety solely from committed actions,” are outside its scope. |
| R13 Preserve output and memory lineage | ACS-Provenance: origin, source_id and derived_from on every data-bearing field, filled by deterministic code (conformance, §7); memory retrieval and store hooks (hooks). Optional profile. | Mechanism. Field-level lineage within a session. Lifecycle controls for persistent state across missions (reuse, quarantine, retirement, restoration) are not addressed. |
| R14 Declare dependency failure behavior | on_decision_failure, default proceed; a separate out-of-band startup policy also defaults to proceeding when the handshake fails; every decision failure recorded as an audit event; a declared liveness mechanism required (§6.4, ACS-Core). | Opposite default. See the next section. |
| R15 Interrupt with accountable continuity | ASK pauses a step and DEFER suspends it pending resolution (§6.4). | Not addressed beyond step-level pausing. Distinct pause, cancel, revoke, quarantine, handoff and recovery effects, with ownership and resumption conditions, remain the deployment’s. |
| R16 Protect decision evidence | A Trace decision event “records the Guardian’s verdict. Only the Observed Agent knows the applied outcome,” and “no decision event proves enforcement on its own” (Trace events); hash-chained SessionContext with an HMAC baseline signature; ACS-Crypto and ACS-Audit as optional strengthening profiles (conformance). | Shared rule and a gap. R16 requires evidence that will “distinguish attempts, authorization, observed execution and outcomes,” and ACS draws the same line between a verdict and an enforced outcome. ACS also states that its failure events are reported by the Observed Agent about its own conduct, so a count of them “is a lower bound,” which parallels R16’s “agent self-attestation alone is insufficient.” The gap: ACS-Core does not assert “that the audit chain is tamper-evident against a compromised Guardian.” R16 makes tamper-evidence and verification by “a verifier other than the acting agent” mandatory, so an ACS deployment that relies on ACS for it needs at least the optional profiles, plus the verification roles and trust anchors R16 asks the profile to identify. |
| R17 Make governed actions auditable | policy_references with an optional policy_version that “SHOULD be populated when replay or ledger-backed policy state matters” (§6.1); OpenTelemetry and OCSF mappings (ACS-Trace, optional). | Mechanism. Reconstruction of what occurred and under which policy. R17 treats isolated policy-decision replay as permitted, not required, and ACS likewise leaves it optional. |
| R18 Govern exceptional operation | No exception-authority construct. | Not addressed. ACS’s audit records of fail-open steps can feed the reporting R18 requires; the bounded, accountable exception authority itself remains the deployment’s. |
The default that matters: failure posture
Under ACS-Core, when no usable decision arrives within the negotiated timeout, the agent applies a declared failure posture, on_decision_failure. The default is proceed, which is fail-open, chosen “so that a slow, erroring, or unreachable Guardian does not halt production.” A deployment may set on_decision_failure: deny instead. The specification is direct about the trade: under fail-open, “an adversary who can disrupt the channel converts control into audit.” It also requires a declared liveness mechanism, because otherwise the chain of defaults “produces silent fail-open on Guardian outage, which the Core baseline cannot allow” (§6.4, ACS-Core). The published specification requires every decision failure to be recorded as an audit event under either posture; the 0.1.4 tag required it for fail-open steps.
R14 takes the other default. Ordinary effects whose required checks cannot be satisfied “shall not proceed,” and continued operation may use only “a pre-authorized bounded continuity mode with declared alternative controls, limits and reconciliation, never an implicit fail-open bypass.” An ACS deployment can configure on_decision_failure: deny and withhold effects whose required checks cannot be met. Continued effects require a pre-authorized bounded continuity mode whose alternative controls, limits and reconciliation satisfy the baseline; setting proceed alone does not establish that. Where a mode cannot meet the baseline, it must be reported as “nonconforming exceptional operation rather than normal conformance.” Reporting that limitation does not turn the mode into conformance. Startup has its own failure path: on_decision_failure is not negotiated when the handshake fails, and the separately configured startup policy defaults to proceeding (§4.1). To satisfy R14, startup must also refuse governed effects or apply baseline-conforming, pre-authorized bounded continuity controls. ACS-Core with either fail-open default left in place, and nothing more declared, does not meet R14. ACS supplies the switch and the audit record; the doctrine says which way the switch has to be set.
ACS already chooses the closed direction in one place. An expired DEFER takes its timeout_decision, which defaults to deny, and the specification explains that this is “deliberately the opposite” of the decision-failure default: a Guardian that flagged a concern and failed to resolve it fails closed, where a silent one fails open (§6).
The Assessment Companion, R19 to R24
ACS is candid that its conformance is self-declared: “Who verifies a conformance claim: nobody, in v0.1.4.” The release ships no conformance test suite and no registry, and ACS-Core conformance does not assert “that a deployment’s policies are strict” (ACS-Core). That is where the Assessment Companion applies.
| Requirement | ACS counterpart | Relationship |
|---|---|---|
| R20 Define and apply an assessment method | No test suite in v0.1; the conformance page tells a buyer to test a deployment “the way you would check any other vendor claim.” | Not addressed. R20’s criterion structure (applicability, evidence, method, expected outcomes, failure rules, retest triggers) can be applied to an ACS deployment. Its permitted-and-challenge cases for R06, R07 and R14 exercise exactly the hook, honoring and failure-posture behavior mapped above. |
| R23 Make precise claims | The repository’s conformance claim template pins the specification “by tag and full commit SHA,” and the 0.1.4 CHANGELOG says to name the release or commit when claiming ACS-Trace conformance. | Shared rule. Both require a claim to name the version it was made against. They differ in subject: an ACS claim is about an implementation, while R23 holds that “only a specified operated deployment can support a deployment claim.” An ACS-Core claim is therefore at most documented-design or examined-configuration evidence toward a deployment claim, not the claim itself. |
| R24 Control profile changes and retests | ACS-Inspect-Dynamic emits agbom/changed on component changes; pre-1.0 patch releases may change the ACS-Core baseline. | Mechanism. A component-change event is a usable trigger for R24’s impact review. A profile that depends on ACS should pin the ACS release and commit it was assessed against, and treat an ACS release as a change to its evidence mechanisms. |
| R19, R21, R22 | None. | Not addressed. Corroboration evidence, assessor disagreement and useful-completion measurement are assessment concerns outside an interface specification. |
Where each is thinner
What ACS supplies that this doctrine does not
Everything on the wire. The doctrine deliberately defines no message format, hook list, verdict vocabulary or error codes. ACS defines all of them, together with a capability handshake, replay protection, a signature baseline, OpenTelemetry and OCSF mappings, and AgBOM serialization, and its repository carries reference implementations. Where a deployment needs a standard way for an agent framework and an external enforcement point to talk, ACS is an open candidate for that layer, and the doctrine has nothing to add to the wire format itself.
What this doctrine supplies that ACS does not
The obligations around the hook and the method for checking them. Of the eighteen baseline requirements, ACS shares a rule with three (R03 within a session, R07 and R16), supplies a mechanism toward eight, defaults the other way on one (R14), and leaves six unaddressed (R04, R08, R10, R12, R15 and R18). Of those, R08 is the one ACS has scheduled for its next version. For the requirements where ACS supplies only a mechanism, the doctrine supplies the rule the mechanism has to be configured to meet, and the Assessment Companion supplies the method and claim discipline that ACS v0.1 leaves to the buyer.
Why complementary, and why the doctrine’s criteria sit above ACS
ACS answers the question of how an agent asks, and how the answer is enforced and recorded. This doctrine answers the question of what a deployment must be able to show within a declared boundary, and how that is checked. A deployment can be ACS-Core conformant and fall short of the baseline, for example by running the default failure posture with no declared continuity mode, or by running a permissive Guardian that ACS itself calls conformant. A deployment can meet the baseline without ACS, using a different enforcement interface. The two do not compete: one supplies an interface, the other sets the obligations and the conformance criteria a deployment using that interface is assessed against.
Both documents refuse to let a capability claim stand in for evidence about a deployment. The doctrine states that a product’s capabilities do not establish conformance of a system operated with it. ACS states that a conformance claim is the implementer’s own assertion until a test suite and the governance to run it exist, which its conformance page says is “scoped to a later release.”
Primary sources
- OWASP Agent Control Standard, project site and documentation (read 14 November 2026; presents release 0.1.4, specification v0.1.0).
- ACS Instrument specification: design principles, architecture, dispositions, honoring decisions and failure posture (§6.4), MODIFY-incapable clients (§6.5), provenance, approver model (§9), deferred scope (§15)
- ACS conformance profiles: ACS-Core, ACS-Trace, ACS-Inspect, ACS-Provenance, ACS-Crypto and ACS-Audit
- ACS Trace events, hook catalog and Inspect (AgBOM)
- ACS repository: README, CHANGELOG at v0.1.4, LICENSING at v0.1.4, and the conformance claim template
- Governed Autonomy Doctrine v4, mandatory runtime baseline and Assessment Companion
Mapped 14 November 2026 against ACS release 0.1.4 (specification v0.1.0) as published. ACS is pre-1.0 and its baseline may change in patch releases; when it does, this page will be revised to match it rather than defended.