xChk against AGSC v1.0.0 (Agentic Governance & Security Controls)
Self-assessment, 6 Nov 2026 (Sean: "rate xchk against AGSC"). Not a conformance claim.
What AGSC is. An open control catalog (CC BY 4.0) from COHUMAIN Labs and SafeAlign AI: 25 controls in four pillars (Safety, Alignment, Governance, Security), each derived from a named MIT AI Risk Repository risk, mapped to MITRE ATLAS / ATT&CK, EU AI Act, NIST AI RMF and ISO/IEC 42001, with "key requirements" and named "evidence". Three cumulative tiers: Foundational (8 controls), Managed (18), Assured (25). Source: github.com/HimJoe/Agentic-governance-and-security-controls-by-COHUMAIN-Labs-and-Safealign-AI (COHUMAIN-AGSC-repo/controls/CONTROLS.md, docs/conformance.md); the vendor page is safealignai.io/resources (PDF behind a lead form; the standard is for agentic and generative AI in general, not scientific AI). SafeAlign sells the governance plane the catalog describes; the catalog itself names no product.
How this is scored. AGSC is written for the deployer of an agent (the organisation), not for a component. xChk is one component: the authorization-and-accountability layer (gate, registry, ceremony, receipt). So each row says what xChk itself satisfies, what it gives the deployer as evidence, and what is not xChk's job.
| Verdict | Meaning |
|---|---|
| MEETS | within the surface a component has, every requirement is held and its evidence is machine-verifiable; a requirement that is the organisation's own act (a signature on minutes, a calendar, host isolation) is named in the last column. No control is met without the organisation: 13 is a component-scope count, not 13 controls a deployer could declare on xChk alone |
| PARTIAL | the technical core is xChk's and a substantive requirement of the control itself is not |
| N/A | outside xChk's scope by design: model lifecycle, generated content, procurement, sustainability |
Five of the 13 are scoped to the surface a component has and say so in the verdict: SAF-02 application layer, ALN-03 agentic surface, ALN-05 the action half, GOV-07 agentic artefacts, SEC-05 own system. (Legend tightened 7 Nov 2026 after the Sourcery review on our PR to the AGSC repo pointed out that qualified MEETS rows strained the earlier wording; no verdict changed.)
Score
| MEETS | PARTIAL | N/A | |
|---|---|---|---|
| Tier 1 Foundational (8) | 7 | 0 | 1 |
| Tier 2 adds (10) | 4 | 4 | 2 |
| Tier 3 adds (7) | 2 | 1 | 4 |
| All 25 | 13 | 5 | 7 |
Read: xChk is a Tier 1 engine. Of the eight Foundational controls it meets seven (ALN-01, ALN-02, SAF-01, SAF-02, SEC-02, SEC-03, SEC-04); the eighth (SAF-03, content safety) is a different product. The seven N/A rows are model lifecycle and content controls, the deployer's by design. First pass on the morning of 6 Nov 2026 scored 8 / 10 / 7; the afternoon's work (see "What changed", below) moved five rows, two of them by building and three by looking harder at what already shipped.
Safety
| Control | Tier | Verdict | xChk evidence | Not xChk's |
|---|---|---|---|---|
| SAF-01 Autonomous System Oversight | T1 | MEETS | Behavioural baseline = the plate's grants (areas × allow/attest/forbid × maxAmount), set by the owner before the first action. Every gated action is a ledger record (plate, tool, target, decision) emitted at the hook, median ~290 ms, under the 1 s bar. Since gate 0.1.33 (released to in.xchk.io 7 Nov 2026) the record also seals the agent's own rationale (Claude Code's Bash description); proven on the real path the same hour: GEN-T9D0 records 6761, 6764 and 6766 each carry the reason the agent gave its command. Seven deviation thresholds already existed, each wired to the owner's phone: rate window, spend window, maxAmount, the runaway ladder, behaviour drift, lockdown, coverage. The review pack (GET /api/agents/:id/review-pack, review-pack.html) is the monthly review: the thresholds as configured, what each tripped, the owner's disposition of each, with a sign-off block against the pack hash. | Signing the minutes. |
| SAF-02 Emergency Stop & Containment | T1 | MEETS (application layer) | The lockdown (MACG-004) was already the control, under-read in the morning pass: one POST, no passkey and no reason between the hand and the brake; four scopes (plate, host, org, registry); the authorize path refuses every action before any grant is read and fails closed if the lockdown state cannot be read; a registry-signed statement the host verifies so a tampered DB cannot forge "not locked"; a halt command to the resident manager for the action already running (category 0 terminates sessions); CAEP session-revoked to Shared Signals; dual-control clear with a reason. Activations are sealed with actor. New 6 Nov 2026: halt_bench.cjs in the on-host benchmark tier measures stop → next action refused per scope against the 5 s SLA, and goes red if an active stop lets anything through or a cleared one still refuses. Run on in.xchk.io 7 Nov 2026: 20/20 halted across the four scopes, median 9 ms, p95 221 ms, every forbid sealed, every cleared stop resumed. | Isolating connectors and processes xChk does not govern. |
| SAF-03 Content Safety & Misuse Prevention | T1 | N/A | The context screen catches injected instructions in tool output (spikee 694/742, 0/300 false positives), not toxicity or regulated-claim content. | Content filters on generative output. Different product. |
| SAF-04 Output Factual Accuracy | T2 | N/A | Groundedness benchmarks. | |
| SAF-05 AI Output Monitoring & Safeguards | T2 | PARTIAL | The HITL half, in full: an attest disposition holds the act for the owner's passkey; the approval record is a signed, anchored receipt with basis and decision blocks (DEMM-Bench 8/8), verifiable offline with no call to the registry. This is the strongest evidence anywhere in the catalog for "approval records on regulated outputs". | Bias / hallucination / drift monitoring of outputs. |
Alignment
| Control | Tier | Verdict | xChk evidence | Not xChk's |
|---|---|---|---|---|
| ALN-01 Autonomy Boundary Control | T1 | MEETS | This is the product. Declarative per-plate policy (grants), deny-by-default, value caps (maxAmount), approval-required class (attest), narrowed per run (grantMask ⊂ grants). Enforced at the gateway, out-of-band: the decision comes from the registry, not from anything in the prompt. "Cannot be overridden by prompt content" is measured, not asserted: bypass benchmark 23 techniques / 79 cases / 0 silent passes, 15 techniques still held under an allow-all grant by self-protection alone, canary proves each bench goes red against a broken gate. The quarterly enforcement test is run-benches.sh, run before every deploy. | Versioned policy file export (grants live in the registry, exportable via the API; a flat file is a format question). |
| ALN-02 Human Oversight & Intervention | T1 | MEETS | Pre-emptive by construction: the hold happens before the act, on the owner's phone, under a passkey; the agent sees a slow tool call. Approval-gate records are the ceremony receipts. Override and escalate exist (owner holds, grants change). | Operator runbook text. |
| ALN-03 Decision Interpretability | T3 | MEETS (agentic surface) | The public receipt (receipt.html, no login) already carries "Why it asked" (reason, code, the ask record) and "The decision" (who, how authenticated, when, any amendment), readable by a third party with no account. The control accepts "evidence the explanation is surfaced in the decision record"; it is. | An explanation written for the affected person in a consumer flow (Art. 13 audience) is the deployer's text. |
| ALN-04 Model Explainability | T3 | PARTIAL | For the agentic surface the control accepts "decision/action trace samples": the ledger and the per-run trust record are exactly that, signed. | Model-level explanation methods. |
| ALN-05 Output Labeling & Action Attribution | T2 | MEETS (action half) | Every act is attributed to the plate and to the accountable human behind it, signed (Ed25519 over JCS), anchored on-chain, and the attribution survives export: the offline verifier checks it with no shared secret. That is attribution that persists downstream, which is the control's second requirement. | Labelling AI-generated content (documents, text). Not an xChk surface. |
Governance
| Control | Tier | Verdict | xChk evidence | Not xChk's |
|---|---|---|---|---|
| GOV-01 Data Provenance & Input Integrity | T2 | PARTIAL | Runtime inputs: the context screen flags and labels untrusted content in tool output before the agent acts on it; every data access through a gated tool is in the audit trail. | Training-data lineage, RAG allow-lists. |
| GOV-02 AI Change Governance | T2 | MEETS | For the control plane itself: gate releases are signed and manifest-hashed, every slot reports its version (GATE_BEHIND_RELEASE), grant changes are logged ceremonies. The review pack's change register (6 Nov 2026) lists every entitlement, key, role, ceiling and install change in the period from the audit log, and flags models and harnesses first seen in the period against the plate's history. | Re-evaluation of the deployer's own prompts before release; the sign-off. |
| GOV-03 Third-Party & Supply-Chain Oversight | T3 | N/A | Contracts. | |
| GOV-04 Model Training Governance | T3 | N/A | ||
| GOV-05 Model Evaluation | T2 | N/A | (xChk evaluates itself: AgentDojo, spikee, AGB, DEMM; that is SEC-06, not GOV-05.) | Evaluating the deployer's model. |
| GOV-06 Drift Management | T3 | N/A | ||
| GOV-07 AI Lifecycle & Registry Management | T3 | MEETS (agentic artefacts) | The registry is a registry: one plate per agent, gate version per install, harness and model per run, census of live hosts (install-fsck). "Which combination was live when the incident occurred" is answerable per run from the trust record, and per period from the review pack (models and harnesses seen, gate versions now, runs with harness and model). | Versioning the deployer's prompts, RAG sources and memory configs. |
| GOV-08 Environmental & Carbon Governance | T3 | N/A | (The run's usage field records tokens; a carbon estimate could be derived, nothing does.) |
Security
| Control | Tier | Verdict | xChk evidence | Not xChk's |
|---|---|---|---|---|
| SEC-01 Multi-Agent Coordination Security | T2 | PARTIAL | One identity per agent, cross-organisation acts governed by grants and level ceilings, maxActiveRuns and run expiry bound runaway loops. Hub work orders record submitter and owner separately. | An interaction map, a documented inter-agent protocol, simulation tests. The STRIDE assessment names delegation protocol as an open gap. |
| SEC-02 Agent Identity & Least-Privilege Authorization | T1 | MEETS | Unique non-shared identity per agent (plate, private_key_jwt, did:web:<registry>:plate:<PLATE>); task-scoped runs that can only narrow (grantMask, expiresAt); HITL for high-impact acts (attest); entitlement review in the console with accountableOwnerEmail per plate. The review pack is the quarterly entitlement review on paper: authority held beside authority exercised, areas acted in with no grant flagged, approvers who are not the accountable owner flagged. | The quarterly review happening is the deployer's calendar. Standing grants do exist (the owner's consent), ceilinged by role and level; AGSC's "no standing broad permissions" is met in spirit (ceiling + narrow-per-run), not by abolishing standing grants. |
| SEC-03 Prompt Injection & Adversarial Input Defense | T1 | MEETS | Trust-context separation is architectural: the authorization decision is made by the registry, outside the model's context, so injected text cannot widen it (the benchmark's 0 silent passes). The screens are an input/output firewall at the tool boundary, measured on spikee (94 %, English 100 %, 0 false positives) and OWASP LLM01 payloads. One stated limit: 48 non-English "repeat N times" misses. | SIEM rules and SOC owners. |
| SEC-04 Observability, Anomaly Detection & IR | T1 | MEETS | Per-call structured traces at the hook; absence of telemetry is an incident: the manager census marks a host MGR_DARK and a run with unaccounted references is flagged (omission evidence, 29 Aug 2026). Detection axis: 14 scenarios, 0 executed-and-undetected, 0 false alarms. Incident and complaint packages are generated from the ledger. | A written AI-incident runbook with severities and SLAs. |
| SEC-05 AI Threat Modeling | T2 | MEETS (own system) | STRIDE self-assessment with named gaps; the bypass benchmark spec is a threat-to-control traceability matrix refreshed on each gate change. | The deployer's threat model of their own agent and connectors. |
| SEC-06 AI Security Testing & Red-Teaming | T2 | MEETS | Adversarial suite in the release pipeline: run-benches.sh + security-scan.sh must be green before deploy; release blocked on red; regression history committed per run; external suites (AgentDojo 25/26, spikee, Murtaza payloads) reproduced. | Red-teaming the deployer's model. |
| SEC-07 Memory & Context Lifecycle Security | T2 | PARTIAL | xChk holds no agent memory, so cross-tenant bleed through xChk has no path. An approval is bound to one action (authRef, paramsHash) and cannot answer a second; the gate's session pin keeps one session on one plate. Retention is stated (DISCLOSURE-READINESS.md §4: sealed records are never pruned, chains are append-only). | Isolation and expiry of the agent's own memory store, and a purge policy for it. Not ours; stays PARTIAL. |
What changed on 6 Nov 2026 (Sean: "can we improve our score" → "do it")
Built, tests green (xchk-server 1763 pass, 0 fail; offline benchmark tier GREEN; security scan GREEN with gitleaks and semgrep SKIPPED on seanair):
- Rationale on the record (SAF-01).
ActionRecord.rationale(≤300 chars) joinsRECORD_CONDITIONAL_KEYS, so it is sealed and recomputed by the console, the incident package and the offline verifiers;js/ledger-verify.jscarries the mirrored key. Server side is in. The gate side is a patch for Sean to apply and release (docs/gate-patches/2026-11-06-rationale.diff: the hook passes Claude Code's Bashdescription,xchk_client.authorizesends it asrationale; applies cleanly to the source, compiled on a scratch copy). The gate never reads it: the decision does not depend on the agent's stated reason. - The review pack (SAF-01, SEC-02, GOV-02, GOV-07).
reviewPackService.buildReviewPack(agent, {from, to}),GET /api/agents/:id/review-pack,review-pack.html+js/review-pack.js, linked from each plate card beside the evidence report. Seven sections: authority held, authority exercised (area × disposition, no-grant areas flagged), who answered (approvers vs the accountable owner), watches / what tripped / disposition, runs, what changed (register + first-seen models and harnesses + gate versions) with emergency stops, coverage and a sign-off block against the pack hash. Reads only records that exist; regenerable for any past period. - The halt bench (SAF-02).
halt_bench.cjs, on-host tier ofrun-benches.sh: per scope, stop → next action refused, latency against the 5 s SLA, sealed forbid present, cleared stop resumes. On in.xchk.io, 7 Nov 2026: 20/20 halted, median 9 ms, p95 221 ms, the whole on-host tier GREEN (the first run read the refusal's wrapper and scored every real halt as a miss; fixed the same hour, 0558314dd). - Three rows re-read, no code: the lockdown already is the SAF-02 control; the public receipt already carries the ALN-03 explanation; approvals are already single-use and retention already stated (SEC-07 facts corrected, rating unchanged).
Deployed 7 Nov 2026 (server a22a9ff on both registries, docroots at 5453fefb3 and after). Gate 0.1.33 released to in.xchk.io by Sean the same night and installed on seanair; rationale arrives on live records. Tag v0.1.33 signed (on starshout, where the key lives) and 0.1.33 published to xchk.me the same night: both registries serve the same manifest hashes (plugin ad8f6270…, standalone 24628c42…) and list the patched hook. Open: bunnymini's copies (0.1.31) when it is reachable.
What AGSC gets right, and the one thing it lacks
It gets the thesis right: for an agent, governance ("what may it do") and security ("what can a compromised one be made to do") are one problem (docs/the-standard.md, "The core argument"). That is the xChk position stated by someone else, and the eight AI-native controls (SAF-01, ALN-01, SEC-01, SEC-03, SEC-07 among them) are the right eight to call new.
What it lacks is an evidence grade. The catalog says conformance is "verifiable, not assertional", but every evidence item is a document the deployer writes about itself: minutes, dashboards, config exports, "approval records". Nothing says who the approver was, how the record is bound to the act it approved, whether a third party can check it without trusting the deployer, or whether a missing record is detectable. ALN-02 and SEC-02 (Tier 1) require "approval records for high-impact actions", SAF-05 (Tier 2) approval records on regulated outputs, and all three accept a log line. An xChk receipt and a print statement in the agent both satisfy the text.
The ask, if we engage (info@cohumain.ai, or the GitHub repo as a PR to docs/conformance.md): add one column to the evidence model, E0 asserted / E1 logged / E2 signed by the actor / E3 signed by the accountable human and independently verifiable, and set E3 as the bar for the approval-record items from the tier each enters at (ALN-02 and SEC-02 at Tier 1, SAF-05 at Tier 2). Opened as RFC #4 and PR #5 on their repo, 7 Nov 2026. The offline verifier and a live receipt are the demonstration. This is the same ask we made of NIST AI 200-2 (evidence contract, custody under Goodhart) and of DEMM; AGSC is the third catalog in 2026 to measure the approval and not the approver.
Disclosure
Nothing new is disclosed here: every mechanism cited is already shipped and public or already logged in docs/NEXT-FILING-PENDING-MATTER.md. The evidence-grade proposal is a property-level ask, not a mechanism.
Source of record: docs/agsc-assessment.md in the xChk repository; this page is its rendered copy.
