
TL;DR
Security advisory ingestion should produce versioned, permission-aware evidence, not an undifferentiated collection of documents. Preserve who published each claim, which product conditions it covers, when it changed, what use is permitted, and which recommendations depend on it. Keep vulnerability intelligence, hardening guidance, scanner coverage, and approved organizational policy distinct.
This installment builds the evidence supply for Example Enterprise: a source registry, a normalized advisory/control contract, and an offline retrieval walkthrough. A withdrawn-fix scenario demonstrates how new evidence changes the supported answer without retraining the model or pretending that stale deployment observations have become current.
The operating principle: update the evidence quickly, but promote authority deliberately.
In This Article
- Introduction
- Choose Sources by the Questions They Can Answer
- Register the Source Before Enabling the Connector
- Build Intake Around a Preserved Source Record
- Normalize Advisories and Controls Without Flattening Them
- Retrieve Applicable Evidence Before Asking for a Recommendation
- Replay a Withdrawn Fix Without Retraining
- Make Outages, Conflicts, and Malicious Content Visible
- Operate the Evidence Supply as a Service
- Conclusion
Introduction
The remediation ticket still recommends Example Renderer 1.4.3. Its source link opens, and the supplier name is correct. The problem is that the supplier has withdrawn that fix since the ticket was drafted.
A knowledge pipeline that preserves only a page address cannot explain what supported the original recommendation. A pipeline that retains the old text but ignores revisions can continue recommending the withdrawn fix. Neither problem is solved by making the model more confident.
The inventory supplies the assets against which external and internal evidence must be matched. The reference design assumes an authorized inventory, an owner for source admission, and separate storage for captured documents and derived search records. The local examples specify an offline implementation using synthetic records. They do not provide a downloadable program or measured ingestion results.
Choose Sources by the Questions They Can Answer
Do not assign one trust score to an entire corpus. Establish authority for a particular claim: a supplier’s supported upgrade path, a catalog’s exploitation signal, or an organization’s approved operating requirement.
Use this source map for the reference implementation:
| Evidence class | Preferred source | What the pipeline must preserve |
|---|---|---|
| Affected products and supported fixes | Supplier advisories and exact-release upgrade documentation | Product branch, prerequisites, exceptions, fix status, and revision |
| Vulnerability identity and enrichment | CVE records and NIST’s National Vulnerability Database | Record publisher, status, product assertions, severity source/version, and corrections |
| Exploitation evidence and forecasts | CISA KEV and FIRST EPSS, stored separately | Signal type, observation or score date, scope, and provenance |
| Hardening and secure-development guidance | Vendor guidance, NIST, OWASP, and appropriately licensed baselines | Control objective, applicable version, publication status, and verification method |
| Organizational requirements | Approved policies, architecture decisions, and runbooks | Approver, effective scope, revision, exceptions, and expiry conditions |
| Detection capability | Documentation and rule metadata for the selected assessment tool | Tool edition/version, check identity, supported target, and authentication requirements |
The CVE Record Format provides structured vulnerability records, while NIST maintains NVD as a repository of software and hardware vulnerability information. In this design, preserve their contributions separately from a supplier’s remediation instructions rather than merging every assertion into one unattributed description. CVE Program: CVE JSON record format
CISA’s Known Exploited Vulnerabilities (KEV) catalog concerns vulnerabilities exploited in the wild. FIRST’s Exploit Prediction Scoring System (EPSS) estimates the probability that a published CVE will be exploited in the wild within the next 30 days. Neither establishes that EE-APP-001 is affected or already compromised. Absence from KEV is not a clean assessment, and a missing EPSS value is not zero probability. CISA: CISA Releases New Known Exploited Vulnerability Nomination Form
CISA’s official repository supplies KEV in JSON and CSV, together with a JSON schema. FIRST recommends daily files for bulk EPSS synchronization and its API for limited lookups. Retain the score date and model version where supplied; an import timestamp should not replace them. Keep a feed’s published due date separate from the organization’s applicable remediation deadline. CISA: cisagov/kev-data
Register the Source Before Enabling the Connector
Make source admission a configuration decision, not something a retrieved document can grant itself. The registry should identify the publisher, approved locator, owner, permitted uses, processing boundary, access rules, expected update behavior, and response to failure.
Review ingestion, redistribution, model-provider processing, and training separately. CIS, for example, publishes product-specific terms and a distinct license link for free Benchmark PDFs. Do not infer unrestricted reuse from download availability. CIS
Define a source-registry.json contract that initially enables only authored local fixtures. Keep real-feed collection disabled until the organization establishes the applicable rights, endpoint, credentials, and operating limits.
This abbreviated entry illustrates the local advisory source. The one-day threshold is a lab choice, not a recommended universal freshness target.
source_id: EE-SRC-ADVISORY record_kind: advisory transport: local_file enabled: true admission_status: synthetic_lab_only source_owner: security-engineering allowed_uses: - local_evidence_retrieval training_allowed: false processing_boundary: EE-LAB allowed_roles: - security-analyst max_successful_check_age_seconds: 86400 on_failure: preserve_history_and_block_current_retrieval
Replace the operating choices through review, not by copying the synthetic approval into production. A registered public source may be appropriate for advisory lookup while still being inappropriate for sending private inventory details to an external service.
Build Intake Around a Preserved Source Record
Separate collection, interpretation, and publication into retrieval. The important boundary in this diagram is that fetching a document does not immediately make it eligible to support an answer.

Constrain Collection and Parsing
For a live connector, allow only reviewed destinations and required protocols. Disable automatic redirects unless the connector validates each destination; apply network restrictions as well as application checks. OWASP’s server-side request forgery guidance explains why an initially permitted URL is not sufficient protection against redirect-based scope changes. OWASP: Server Side Request Forgery Prevention Cheat Sheet
Give collectors no deployment credentials. Bound response size, decompression, execution time, retries, and parser resources. Treat a login page returned with a successful HTTP status as a failed collection, not a new advisory. For paginated feeds, retain the window and cursor, and advance the checkpoint only after the required pages have been committed successfully.
Prefer structured data when available. The OASIS Common Security Advisory Framework (CSAF) 2.0 standard defines product, vulnerability, remediation, and revision-history structures. Validate against the format version the supplier actually emits. The local contract illustrated here is not a CSAF implementation. OASIS CSAF 2.0
For human-readable guidance, preserve tables, headings, footnotes, and the conditions attached to a recommendation. A parser that extracts a target version but drops the required intermediate upgrade has changed the meaning. Hold incomplete extraction for review rather than letting the model supply the missing step.
Preserve the Original and the Interpretation
Store a permitted source snapshot with publisher identity, canonical locator, source publication/update dates, retrieval time, content digest, and parser version. Keep the normalized record beside it, not in place of it. Preserve date precision; do not invent a publication time or time zone that the source never supplied.
A content hash identifies the bytes captured. It does not authenticate the publisher or prove that the content is correct. OWASP’s RAG security guidance also warns that an attacker who can replace both a document and its stored hash can defeat that comparison. Protect the manifest separately; verify publisher signatures where the source actually supplies a supported mechanism. OWASP: Retrieval-Augmented Generation (RAG) Security Cheat Sheet
Deduplicate repeated delivery within the same source, record, and revision. Do not erase independent provenance because two publishers discuss the same CVE. If a page changes without a declared revision change, preserve the new capture and reconcile the discrepancy instead of silently overwriting the old record.
Normalize Advisories and Controls Without Flattening Them
Use a common provenance envelope with type-specific content. Define a security-evidence.schema.json contract that distinguishes an advisory, baseline guidance, and an internal policy proposal.
| Record area | Minimum useful content |
|---|---|
| Identity and history | Source ID, source record ID, revision, document type, and lifecycle status |
| Evidence receipt | Captured-file reference, content hash, parser revision, source dates, and retrieval time |
| Admission and access | Admission decision, environment boundary, permitted audience, and processing restriction |
| Advisory payload | Product/version assertions, configuration prerequisites, remediation state, and detection-coverage status |
| Control payload | Control objective, applicability, verification method, and separate adoption state |
| Supporting passages | Exact text, location within the captured source, and revision-specific identifiers |
Retain product-specific version semantics. An unresolved package or appliance mapping should remain unresolved, not become a broad version range inferred from similar names. Distinguish “no fix supplied,” “fix withdrawn,” and “fix supported”; do not represent all three as a missing text field.
For controls, retain publication status independently from organizational adoption. On November 18, 2026, NIST’s publication listing identifies SSDF 1.1 as final and SSDF 1.2 as a draft. A newer draft can inform review without silently replacing the organization’s accepted baseline. NIST: Secure Software Development Framework, Publications
The synthetic LAB-CTRL-001 recommends restricting preview-service outbound connections. The separate EE-POL-001 proposes adopting that control but has no approval or effective date. It can explain a proposed requirement; it cannot authorize enforcement. Likewise, importing a scanner rule description does not install or validate that rule.
Retrieve Applicable Evidence Before Asking for a Recommendation
Start retrieval with the asset, environment, task, and authenticated caller context. Resolve product identifiers and source revisions through ordinary code. Use semantic retrieval to locate explanatory passages within the permitted candidate set, not to replace exact product or release-branch matching.
For the proposed service, enforce permissions before restricted text reaches a reranker, model, or unapproved embedding provider. Recheck current admission and revision state when assembling the evidence bundle. A vector index returning an obsolete chunk must not override the authoritative current-revision record.
OWASP’s RAG guidance recommends access metadata on chunks, tenant isolation, and permission-aware caching. Apply those controls to derived text, embeddings, citations, and cached answers, not only to original documents. Permission or source changes should invalidate dependent caches; an inaccessible current revision must not cause a silent fallback to an older, accessible answer. OWASP: Retrieval-Augmented Generation (RAG) Security Cheat Sheet
Return an evidence bundle containing the relevant assertions, prerequisites, source passages, dates, and unresolved conflicts. Let the application assign citation identifiers. A model-generated link is not proof that the passage exists or supports the recommendation.
For each consequential proposal, retain the exact evidence revisions it used. That reverse relationship makes a correction actionable: the service can identify which pending proposals require revalidation rather than merely refreshing its search index.
Replay a Withdrawn Fix Without Retraining
Use the existing EE-APP-001 fixture unchanged. Its historical observation reports Example Renderer 1.4.2 on EE-DEP-001, but the current EE-DEP-002 artifact and configuration remain unresolved. The new knowledge pipeline must not manufacture those missing observations.
Admit the Original Advisory
The original synthetic LAB-ADV-001 revision 1 describes version 1.4.2 as affected when the preview condition is enabled and names 1.4.3 as the supported fix. Normalize the document, preserve its passages, and admit its revision through the fixture workflow.
The retrieval result may identify 1.4.3 as the supplier’s stated target. Local applicability remains unknown because the current deployment has not been established. That distinction belongs in the evidence bundle for EE-FIND-001, under the existing EE-EVT-001 scenario correlation.
Admit the Correction in a Separate Test Branch
The withdrawn-fix branch reuses the counterfactual introduced earlier in the series: revision 2 withdraws 1.4.3 without changing the affected conditions or naming a replacement. The original baseline remains available in its separate branch.
Preserve both revisions. Promote the new revision only against the expected current head, preventing a delayed worker from overwriting a newer decision. Rebuild affected retrieval records and invalidate dependent answer caches.
This abbreviated projection shows the revised evidence state:
record_id: LAB-ADV-001
record_kind: advisory
branch: withdrawn-fix
source:
source_id: EE-SRC-ADVISORY
revision: "2"
status: corrected
payload:
product: Example Renderer
affected_versions: ["1.4.2"]
remediation:
state: withdrawn
target_version: null
The advisory is corrected; the proposed fix is withdrawn. Those are not the same as withdrawing the entire vulnerability record.
The supported target becomes null. Local applicability remains unknown. A pending EE-CHG-001 based on the old fix requires revalidation, and EE-VER-001 cannot establish closure. The workflow must not invent 1.4.4 or convert “no replacement fix” into “not affected.”
Only evidence and its derived retrieval state need to change. No model weights need updating. An offline implementation can isolate this behavior without making a model call, so the result cannot be mistaken for improved model reasoning.
Validate the Local Revision Transition
Implement the revision transition using the two synthetic records shown above. First retrieve revision 1, then admit revision 2 and repeat the same query. Verify that current retrieval excludes the withdrawn fix while retaining the historical record and dependent proposal for review.
The first command prints the before-and-after evidence summaries. The second exercises revision handling, duplicate delivery, branch separation, access filters, stale source checks, quarantine, and record-type boundaries. Neither writes production state.
Inspect the transition from target 1.4.3 to null, the unchanged unknown applicability, and the two preserved revisions. The included source hashes are computed from the authored fixture files; they are not supplier signatures or deployment-image digests. The local access checks use synthetic caller attributes and do not validate an identity provider or production authorization service.
Make Outages, Conflicts, and Malicious Content Visible
A current knowledge service needs more than a successful fetch counter. Define what the system does when the evidence cannot support the requested decision.
| Condition | Required behavior in this design |
|---|---|
| Source unavailable or freshness limit exceeded | Preserve history, mark currentness unavailable, and block unsupported claims of a current fix |
| Supplier and aggregator disagree | Preserve both assertions and their scopes; request product-specific resolution |
| New advisory has no supported scanner check | Record the detection gap; continue supported applicability review without inventing a test |
| Current revision is quarantined or access is revoked | Remove it from usable retrieval and invalidate derivatives; do not silently revive its predecessor |
| Retrieved text requests a tool action or policy change | Treat the request as document content; require the independent tool or policy authority |
A source outage should not erase an established exposure or reset its urgency. An owner may continue containment or evaluate a last-known supported action through the existing process, with the missing currentness evidence made explicit. Denying a claim of freshness is different from stopping every defensive action.
Preserve unresolved contradictions without averaging them. An older vendor package can legitimately differ from an upstream version description, while a newer vendor correction can invalidate an aggregator’s fix recommendation. Compare product scope, packaging, conditions, and revision history before deciding which assertion governs.
OWASP identifies indirect prompt injection through external content and cautions that RAG and fine-tuning do not eliminate it. Screening and quarantine are useful, but a missed instruction must still be unable to grant credentials, enlarge scope, or deploy a change. OWASP Gen AI Security Project
Test deterministic handling of records already marked quarantined separately from detection of malicious content. Passing that test would not demonstrate that a model or detector recognizes every malicious document. The later agent-protection installment addresses that separate evaluation problem.
Operate the Evidence Supply as a Service
Assign the platform team responsibility for collection, parsing, storage, and retrieval health. Security engineering owns source authority and interpretation rules. Governance owners approve baseline adoption. Application owners resolve product applicability and operational constraints.
Track the last successful source check separately from the latest published revision and the latest admitted revision. An unchanged, successfully checked advisory is not automatically stale because it was first published months ago. A recently downloaded old snapshot is not automatically current.
Measure publication-to-admission delay where publication time is known, correction-to-dependent-proposal invalidation, failed collection age, unresolved conflicts, and sources missing an accountable owner. Track outstanding gaps alongside completed work; fetched-document counts alone do not show whether relevant guidance reaches the right decision.
Keep scheduled reconciliation alongside change events. On recovery, restore source receipts and revision history, then reapply current revocations, withdrawals, and access decisions before reopening retrieval. Rebuilding a search index must not resurrect guidance that was deliberately removed from active use.
Conclusion
Keeping a security agent current is an evidence-management responsibility. The useful output is an applicable, permitted, revision-specific record that can support a decision and be revisited when its source changes.
Start with a small registry, one advisory path, one baseline path, and the existing inventory. Prove correction handling and visible failure states before expanding ingestion. Keep policy, scanner code, and training examples on their own reviewed promotion paths.
The next installment, Fine-Tune Your Own Security Model on Reviewed Findings and Remediations, uses this evidence foundation to build task examples and compare model adaptation against an unchanged baseline.
Before accepting an agent’s recommendation, ask which source revision supports it and what would cause that recommendation to be withdrawn.
Related Reading
- Fine-Tune Your Own Security Model on Reviewed Findings and Remediations
- Test Your Security Model Before Giving It Operational Access
- Trigger Security Reassessment After Releases, Patches, and Configuration Changes
External References
- CISA: cisagov/kev-data
- CISA: CISA Releases New Known Exploited Vulnerability Nomination Form
- FIRST: Exploit Prediction Scoring System (EPSS)
- FIRST: Get the Data
- CVE Program: CVE JSON record format
- NIST: National Vulnerability Database
- OASIS: Common Security Advisory Framework Version 2.0
- NIST: Secure Software Development Framework, Publications
- Center for Internet Security: Terms and Conditions
- OWASP: Server Side Request Forgery Prevention Cheat Sheet
- OWASP: Retrieval-Augmented Generation (RAG) Security Cheat Sheet
- OWASP: LLM01:2025 Prompt Injection
1 thought on “Keep Security Agents Current with Trusted Advisories and Best Practices”