
TL;DR
Event-driven security reassessment should connect changes to the evidence and assessments they invalidate. Internal deployments change what may be running. External advisories change what is known about software already deployed. Both should create bounded, owned work against the same inventory, without automatically authorizing a scan, patch, or production change.
Use a durable event inbox, version-specific assessment requirements, and scheduled reconciliation. Deduplicate repeated delivery, but do not mistake a processed notification for completed assessment. This installment provides the event/action matrix and an offline advisory-routing example that preserves the shared lab’s unresolved deployment evidence.
Trigger the decision when relevant facts change. Keep the work open until its required evidence exists.
In This Article
- Introduction
- Route Changes to Decisions Before Routing Them to Scanners
- Separate Event Receipt from Assessment Fulfillment
- Admit Notifications Without Importing Their Authority
- Match the Change to the Exact Evidence Basis
- Deduplicate Delivery Without Suppressing Necessary Work
- Handle Reordered Events Without Reversing Current State
- Route the Supplier Patch Before a Scanner Check Exists
- Recover Delivery Failures and Missed Work Separately
- Validate the Advisory-Routing Contract
- Operate the Delays and the Gaps
- Conclusion
Introduction
The document portal has no new build to assess. Its team has made no further deployment. A supplier publishes a security patch for a component associated with the service, but the release pipeline remains quiet.
The relevant change happened outside the enterprise.
Waiting for the next internal release would leave that disclosure disconnected from the deployed environment. Immediately launching every scanner would create a different problem: more activity without first establishing applicability, supported detection, or operating permission.
This article connects the inventory, knowledge store, and controlled assessment paths developed earlier in the series. The reference organization remains Example Enterprise. The proposed integration uses authenticated source adapters. The local scenario specifies synthetic notifications and SQLite state, rather than supplying a downloadable router or measured live-ingestion, scanner, model, or deployment results.
Route Changes to Decisions Before Routing Them to Scanners
Treat an event as a reason to reconsider something, not as a command. A supplier release first needs a security-relevance decision. A newly observed deployment needs identity and configuration verification. An exploitation signal may change urgency without requiring another identical package scan.
The following matrix defines the proposed routing policy. Every execution path remains subject to the previously established controller and approved assessment profile.
| Trigger | Bind work to | Assessment or decision to create |
|---|---|---|
| Pull request | Tested commit and comparison base | Changed-code, secret, dependency, and infrastructure-definition checks |
| Merge or release-branch update | Exact resulting commit | Appropriate full/baseline assessment as well as change-focused checks |
| New build | Build identity and immutable artifact | Component inventory, software bill of materials (SBOM), and supported image/package checks |
| Staging deployment | Observed artifact, configuration, environment, and test role | Scoped authenticated application and API tests |
| Production deployment | Deployed artifact/configuration and target incarnation | Independent state verification and approved production checks |
| Supplier release or patch | Supplier record, revision, and product branch | Establish security relevance, supported fix status, candidate assets, and owners |
| New, corrected, or withdrawn advisory | Admitted advisory revision and current inventory evidence | Reassess applicability and revalidate dependent proposals |
| New KEV entry or credible exploitation evidence | Vulnerability identity, source, and signal date | Refresh exposure priority and determine whether investigation is warranted |
| EPSS change | Vulnerability, score date, and model version where available | Update the likelihood signal and review queue position when policy warrants |
| Configuration or identity drift | Changed configuration/grants and dependent relationships | Reassess affected controls, privileges, and exposed paths |
| Scanner rule or database update | Admitted check revision and relevant artifacts/targets | Reassess changed coverage and revisit missing-check gaps |
| New hardening guidance or internal policy | Source revision and separate adoption state | Review baseline changes, then assess under approved requirements |
| Scheduled reconciliation | Authoritative snapshots and freshness requirements | Find missed events, stale evidence, incomplete routing, and failed work |
CISA’s official KEV repository supplies catalog data and revision history through JSON, CSV, and repository commits. Use changes to that evidence as inputs to local decisions, not as proof that a particular asset was compromised. FIRST defines EPSS as the probability that a published CVE will be exploited in the wild within the next 30 days. Its changing estimate is a prioritization signal, not a target-specific assessment. CISA: cisagov/kev-data
Keep those signal types distinct. An EPSS increase should not silently change unknown applicability to affected. Nor should a new release number become a supported security fix before the supplier’s guidance establishes that relationship.
Separate Event Receipt from Assessment Fulfillment
Maintain two records: what arrived and what must now be established. The event inbox answers whether a notification was accepted. The assessment-obligation ledger records the specific evidence or review the event requires.
An obligation is a requirement to obtain evidence, not authorization to execute a tool. It can remain blocked by stale inventory, unavailable checks, missing ownership, or an unapproved operating window after the notification has been fully processed.
Notice where scheduled reconciliation enters the architecture. It reconstructs required work from authoritative state rather than depending on another copy of the original notification.

Use deterministic code for established product mappings and routing rules. A model can explain an ambiguous advisory or draft an owner-facing summary after permitted evidence is assembled. It does not need to decide whether a deployment event belongs to the deployment workflow.
Admit Notifications Without Importing Their Authority
Normalize provider-specific notifications into a small internal envelope. CloudEvents 1.0.2 defines source plus id as the identity of a distinct event and permits repeated delivery to retain that identity. Its serialization uses specversion: "1.0". The envelope does not authenticate the producer or authorize its requested consequences. CloudEvents: CloudEvents – Version 1.0.2
Use these fields for an authored knowledge-store notification, stored as fixtures/advisory-event.json in an implementation:
{
"specversion": "1.0",
"id": "EE-MSG-ADV-001",
"source": "urn:example:security:knowledge",
"type": "example.security.advisory.changed",
"subject": "LAB-ADV-001",
"time": "2026-11-18T14:10:00Z",
"datacontenttype": "application/json",
"data": {
"record_id": "LAB-ADV-001",
"revision": "1"
}
}
The timestamp is synthetic. The source identifies the internal knowledge adapter, not a supplier whose identity the payload can assert. Keep EE-EVT-001 as the existing scenario correlation; each notification receives its own EE-MSG identifier.
A real receiver must map authenticated integration identity to allowed event types and organizational scope. Derive tenant and environment from that binding rather than accepting arbitrary payload fields. References resolve through registered stores; incoming messages must not supply executable commands, credentials, or arbitrary fetch destinations.
Verify the Provider’s Actual Signed Material
GitHub documents HMAC-SHA256 verification through X-Hub-Signature-256, computed over the received payload and checked with constant-time comparison. Verify before parsing or rewriting the body, and retain a protected receipt separately from normalized content. GitHub: Validating webhook deliveries
Do not assume the signature covers every transport header. OWASP’s Webhook Security guidance specifically distinguishes GitHub’s body-only signature from schemes that also authenticate a delivery timestamp and event ID. A delivery identifier helps with normal redelivery, but does not independently provide authenticated replay protection. OWASP: Webhook Security Cheat Sheet
Apply provider-specific freshness and replay controls where supported, plus downstream idempotency. A signed notification remains source data. It cannot grant itself production access or approve the policy changes it describes.
Match the Change to the Exact Evidence Basis
For internal changes, bind the requirement to the exact commit, build, artifact digest, configuration revision, deployment incarnation, or permission set involved. Branch names, movable image tags, and reusable hostnames are insufficient substitutes.
For external disclosures, use a reverse component index: advisory product conditions to relevant components, their artifacts, deployments, and owners. Preserve supplier packaging, release branches, backports, configuration prerequisites, and embedded-component relationships. Route ambiguous mappings for review rather than broadening a version comparison until it matches.
Missing inventory deserves its own work item. No product match in an incomplete inventory is not a defensible not-affected decision. Historical evidence can identify a candidate owner while still being inadequate to establish current exposure.
Define the assessment’s basis from all decision-relevant inputs. In the reference design, that includes target identity, inventory evidence, admitted advisory revision, assessment profile, relevant check state, and routing policy. A later disclosure can therefore require new work even when the artifact has not changed.
Before dispatch and before accepting completion, compare that basis with current state. Preserve a result for the artifact actually assessed; do not attach it to a replacement deployment. During a rolling release, retain requirements for old and new instances while both remain in scope.
Deduplicate Delivery Without Suppressing Necessary Work
Use separate identities for notification delivery, desired assessment work, and execution attempts. Repeated delivery should return the existing receipt. Different events describing the same still-current requirement can share one obligation while retaining every causal reference.
Fingerprint the input basis separately from notification identity. Commit the inbox entry, obligation, and causal association together. Reusing a source/event ID with changed content must raise a conflict instead of silently replacing the original record.
A production handoff to a broker introduces another boundary. AWS’s transactional-outbox guidance describes committing a database update and notification intent in one transaction, then publishing from that durable record. It also requires duplicate-tolerant consumers. An outbox does not make the remote assessment exactly once. AWS Documentation
Do not use a permanent deduplication key that prevents future assessment of the same artifact. Changed knowledge, changed checks, or an expired freshness requirement must be able to create a new obligation. Conversely, assigning a new random work ID to every retry defeats deduplication.
Persist the desired work before acknowledging its handoff. Recheck authorization at execution, and use the existing job controller to reconcile uncertain scanner acceptance. Redelivering a notification is not the same as restarting a possibly active scan.
Handle Reordered Events Without Reversing Current State
Treat the event’s revision as a reference to an occurrence, not an instruction to overwrite the current record. Source adapters should establish authoritative ordering through documented sequences or revision relationships. Do not order unrelated revision strings lexically or use arrival time as the product timeline.
Where ordering is uncertain, refresh the authoritative source or retain a conflict. A delayed deployment-completed message must not make an old image current again. A delayed advisory must not restore a withdrawn fix.
For a first offline harness, supply an admitted current snapshot and let notifications trigger planning against it. That isolates routing behavior; source-version admission and distributed concurrency still require separate implementation and validation.
Preserve old occurrences and their consequences. A short-lived vulnerable deployment may no longer be an active assessment target but can still matter to incident review. Superseding its current-state record should not erase the historical exposure interval.
Route the Supplier Patch Before a Scanner Check Exists
Return to LAB-ADV-001 and EE-APP-001. The original synthetic advisory describes Example Renderer 1.4.2 under the preview condition and names 1.4.3 as the supported fix. The inherited inventory still contains historical evidence for EE-DEP-001, while EE-DEP-002’s artifact and configuration remain unresolved.
An admitted advisory notification reaches the router. Its historical product match identifies document-services as the owner. The router creates a deployment-applicability obligation and records two distinct gaps: current deployment evidence is unestablished, and the synthetic check catalog has no relevant scanner check.
The first next action is to obtain fresh component and configuration evidence through an approved method. Security engineering separately tracks detection availability and supported manual validation. The owner can evaluate supplier-supported exposure reduction without waiting for a scanner feed, but any operational action still follows its own approval path.
If fresh evidence later establishes the affected conditions, applicability can become affected even while automated assessment remains unavailable. This installment does not supply that observation, so applicability stays unknown.
When an admitted check update arrives, route the affected requirement again. A new check resolves a method gap; it does not establish the deployed version, authenticate to the target, or prove the check ran. Avoid launching the whole fleet simply because a database revision changed. Use relevant check/component mappings where available; where they are unavailable, record the limitation and choose a bounded broader reassessment.
Preserve the Withdrawn-Fix Counterfactual
The separate withdrawn-fix branch advances LAB-ADV-001 to revision 2 and removes 1.4.3 as a supported target. The router retains revision-1 history but creates work against the admitted revision-2 evidence. A delayed revision-1 notification does not restore the old target.
Any pending EE-CHG-001 proposal depending on that target requires revalidation by the change workflow. The router does not modify an external ticket. EE-FIND-001 remains unresolved, and EE-VER-001 has no basis for closure.
Recover Delivery Failures and Missed Work Separately
GitHub states that it does not automatically redeliver failed webhooks; consumers need their own recovery process. AWS likewise distinguishes best-effort and durable delivery attempts from individual services into EventBridge. A healthy downstream queue cannot recover an occurrence that never reached it. GitHub: Handling failed webhook deliveries
Give every connector a recovery contract: authenticated intake, durable acceptance, bounded retry, source checkpointing, and reconciliation. For polling connectors, retain the last completed snapshot or cursor, overlap update windows where necessary, and commit the checkpoint only after required pages are durably processed. An empty partial response must not advance the source as though enumeration completed.
Retry transient processing failures with bounded backoff and jitter. Quarantine invalid signatures and schemas rather than repeatedly processing them. Put exhausted operational failures in an access-controlled dead-letter queue, meaning a retained failed-message queue, with an owner and a replay procedure. Replay preserves the original occurrence and records a new delivery attempt; it does not manufacture a new change.
At reconciliation, compare current required assessments with valid completion evidence, not just inbox receipts. Recreate missing work, reopen expired evidence requirements, and reconcile failed or uncertain jobs. Assign disconnected assets and unavailable methods visible gaps rather than silently removing them from scope.
Apply per-source intake limits and per-service work budgets. Coalesce redundant requests without delaying advisory withdrawals or revocations behind a long batching window. Revalidate permissions before releasing queued work after a maintenance freeze or outage.
Validate the Advisory-Routing Contract
Implement an offline advisory-routing harness before enabling live intake. Python with standard-library SQLite is one suitable choice. Keep authenticated production identity, source admission, and dispatch outside that initial fixture harness.
Begin with four notification families: advisory changes, deployment observations, check-catalog changes, and reconciliation requests. Treat the other matrix rows as integration requirements. Injected principal identities and snapshots test routing logic, not authenticated production inputs.
Use the following expected outcomes as acceptance cases. Execute the completed router and retain its actual results before reporting successful tests:
| Scenario | Expected routing result |
|---|---|
| Original advisory notification | One owned obligation; deployment and check gaps retained |
| Identical repeated notification | No additional obligation |
| Advisory notification never received | Independent reconciliation creates the missing obligation |
| Old notification after revision-2 admission | No restoration of the withdrawn fix and no additional equivalent obligation |
Also test transaction rollback, persistence across a database reopen, conflicting payloads, cross-boundary rejection, and renewed requirements. Those tests cannot establish webhook security, broker behavior, source completeness, scanner detection, or production authorization without exercising those actual integrations.
Checking whether desired routing records exist is only the first step. A fulfillment reconciler must also inspect current job outcomes and accepted evidence. A successful routing process must not imply that the application passed assessment. This scenario starts no scan or deployment, and current applicability remains unknown.
Operate the Delays and the Gaps
Assign connector and queue health to the platform team, advisory admission and check mapping to security engineering, and local evidence collection to the asset owner. Missing ownership enters an assigned triage queue. Dead-letter records and blocked obligations need owners too.
Measure source-publication-to-admission delay when the publication time is known, admission-to-applicability decision, and decision-to-owner acceptance. Track the age of missing deployment evidence, unavailable checks, failed delivery, and incomplete obligations. Count unresolved cases as well as completed ones; otherwise the slowest work disappears from the performance report.
After recovery, restore receipt history and obligation state, refresh current source and policy decisions, and reconcile before dispatch resumes. The practical acceptance test is to interrupt the path deliberately: lose a notification, repeat it, delay an older revision, and fail a worker. Each condition should leave an explainable requirement or outcome, not a misleading clean result.
Conclusion
Security reassessment should follow changes in deployed state, security knowledge, and assessment capability. An internal release is only one of those triggers. A supplier disclosure can create important work against an unchanged application.
Begin with one service, one trusted advisory source, and a durable record of the evidence each event requires. Add periodic reconciliation before relying on event delivery. Expand execution only through the established policy and job boundary.
The next installment, Prioritize Vulnerabilities Using Exposure, Exploitation, and Business Impact, uses these evidence-backed requirements to decide which work deserves attention first.
The event starts the obligation. Only evidence can satisfy it.
Related Reading
- Scan Source Code, Dependencies, Containers, and Infrastructure as Code
- Build Security Agents with Controlled Access to Scanning Tools
- Keep Security Agents Current with Trusted Advisories and Best Practices
External References
- CloudEvents: CloudEvents – Version 1.0.2
- GitHub: Validating webhook deliveries
- GitHub: Handling failed webhook deliveries
- OWASP: Webhook Security Cheat Sheet
- AWS Prescriptive Guidance: Transactional outbox pattern
- Amazon EventBridge: Delivery level for AWS service events
- CISA: cisagov/kev-data
- FIRST: Exploit Prediction Scoring System (EPSS)