Vulnerability Management
Overview
Vulnerability management is a continuous loop, not a quarterly scan: discover → prioritize → remediate → verify → measure → improve. Every step must produce a durable artifact (a finding with an owner, a ticket, a patch, a metric) or it did not happen.
The core discipline is triage economics. Scanners produce far more findings than any team can fix, and most are never exploited. Spend remediation capacity on the small set that is reachable, exploitable, and material to the business — then prove the rest was consciously accepted.
When to Use
- Standing up a vulnerability management program from scratch
- Designing or tuning a scanning strategy (internal, external, agent, DAST/SAST)
- Building a risk-based prioritization model for a noisy scanner backlog
- Automating patch deployment and remediation verification
- Handling an actively exploited vulnerability or zero-day
- Setting up CVE / vendor advisory monitoring and alerting
- Defining vulnerability SLAs, MTTR targets, or executive reporting
- Integrating scanner output with SIEM, SOAR, ticketing, or a vulnerability database
- Preparing audit evidence (PCI-DSS 11.3, SOC 2 CC7, ISO 27001 A.8.8)
Program Structure
Six named components. If any is missing, the program degrades into "scan and email a PDF."
| Component | Owner | Output | Cadence |
|---|---|---|---|
| Asset inventory | IT / Platform | Assets with owner, criticality, exposure | Continuous |
| Discovery & scanning | Security engineering | Findings with evidence | Daily–weekly |
| Prioritization | Security + asset owners | Ranked remediation queue | Per scan cycle |
| Remediation | Asset owners | Patches, config changes, controls | Per SLA |
| Verification | Security engineering | Re-scan evidence / closure record | Post-fix |
| Measurement | Program owner | Metrics, trends, exceptions | Monthly / quarterly |
Inventory first. Every asset needs a unique ID, an owning team (not an individual), business criticality (Tier 1 revenue / Tier 2 internal / Tier 3 dev), exposure (internet, partner, internal, air-gapped), data classification, OS and software versions, and last successful scan. Assets without an owner are the largest single source of unpatched criticals — route them to a default owner and escalate.
Scope and exclusions. Write down what is out of scope and why: medical devices under vendor contracts, OT with change-control windows, end-of-life systems awaiting decommission. Every exclusion needs an owner and an expiry date.
SLAs. Baseline thresholds; adjust for regulatory context (PCI-DSS requires critical patches within 30 days):
| Severity | Internet-facing | Internal Tier 1 | Tier 2/3 |
|---|---|---|---|
| Critical (CVSS ≥ 9.0 or KEV) | 48–72 h | 7 days | 30 days |
| High (7.0–8.9) | 7 days | 30 days | 60 days |
| Medium (4.0–6.9) | 30 days | 90 days | 180 days |
| Low (0.1–3.9) | 90 days | 180 days | Best effort |
Overrides that beat the table: KEV listing → Critical regardless of CVSS; EPSS ≥ 0.10 with a reachable path → bump one tier; public exploit or active exploitation → emergency change; documented compensating control → may extend, but only via an approved, time-boxed exception.
Exceptions. Each record holds finding ID, justification, compensating control, risk owner, approval and expiry dates (max 90 days, renewable). Review monthly; expired exceptions auto-reopen the finding.
Scanning Strategy
No single scan type covers an environment. Combine them and de-duplicate.
Authenticated (Credentialed)
Default for servers and workstations: the scanner logs in and reads installed packages, registry, and configuration. Higher fidelity, fewer false positives than network scanning.
- Use a least-privilege read-only scan account; store credentials in a secrets manager, never in scanner config
- Verify credentials work — a silently failing credentialed scan degrades to unauthenticated and quietly loses coverage
- Alert when a host's credentialed scan fails two cycles in a row
Agent-Based
Preferred for roaming laptops, ephemeral cloud instances, and networks that block inbound scan traffic.
- Continuous reporting enables near-real-time detection
- Watch agent health: offline agents are invisible assets
- Prevent non-admin users from disabling the agent
- Treat the agent as privileged software — patch and monitor it too
Unauthenticated / Network
Port and service discovery, banner grabbing, TLS configuration. Use for asset discovery and to see what an attacker sees. Expect false positives; treat results as leads, not findings.
External / Perimeter
Scan from the public internet against owned ranges and domains — this is what regulators and attackers see.
- Confirm authorization (cloud provider approval, ISP notification, written scope)
- Maintain a live inventory of external IPs and DNS names; shadow-IT domains are the most common external exposure
- Weekly minimum, daily for Tier 1 internet-facing assets
- Include certificate expiry and weak-cipher checks
DAST
Test running applications: OWASP ZAP, Burp, Nuclei. Run in CI against staging and continuously against production with safe policies.
- Authenticate the scanner — most high-impact bugs live behind auth
- Avoid destructive payloads in production (no DELETE, no mass form submission)
- DAST output needs human confirmation; scanner "critical" often means "reflected parameter exists"
SAST and SCA
- SAST (Semgrep, CodeQL, SonarQube): run per pull request; block only on high-confidence critical rules, or developers learn to ignore it
- SCA (Dependabot, Renovate, Trivy, Grype, OWASP Dependency-Check): required per build; emit an SBOM per release
- Feed SAST/SCA results into the same prioritization queue so owners see one list, not four dashboards
Container and Cloud
- Scan images at build time and continuously in the registry; block deploys when a patched base image exists
- Scan running workloads for drift from the built image
- Cloud configuration findings (CSPM) belong in the same program — a public bucket is a vulnerability in business terms
Scheduling
Daily for internet-facing and Tier 1; weekly for internal Tier 2; monthly for Tier 3 and developer endpoints. Scan after major change windows and immediately after patch deployment for verification. Stagger windows to avoid saturating links or tripping IDS alarms during business hours.
CVSS Scoring
CVSS produces a severity score, not a business priority. Know which vector you are reading.
Base (Intrinsic, Static)
- Attack Vector (AV): Network / Adjacent / Local / Physical
- Attack Complexity (AC): Low / High
- Privileges Required (PR): None / Low / High
- User Interaction (UI): None / Required
- Scope (S): Unchanged / Changed
- C/I/A impact: High / Low / None
v3.1 and v4.0 differ in mechanics (v4.0 adds Attack Requirements and splits subsequent-system impact) but the decision logic is the same. Prefer the version the vendor published; do not re-score to make vendors comparable.
Bands: 0.0 None · 0.1–3.9 Low · 4.0–6.9 Medium · 7.0–8.9 High · 9.0–10.0 Critical.
Temporal (Exploit Maturity)
- Exploit Code Maturity (E): Unproven / PoC / Functional / High
- Remediation Level (RL): Official Fix / Temporary / Workaround / Unavailable
- Report Confidence (RC): Unknown / Reasonable / Confirmed
This is where most programs should spend their energy: a 9.8 with no exploit and no reachable path matters less than a 6.5 with a working public exploit.
Environmental (Your Context)
- Modified base metrics: is the component present and configured vulnerably?
- C/I/A requirements (CR/IR/AR): how much does this asset matter?
Environmental scoring is manual, so apply it only to Tier 1 assets and to any finding whose default severity would trigger an emergency change.
Discipline
- Never average CVSS scores into a risk rating; use the maximum plus counts
- Record the vector string, not just the number — the number is unauditable alone
- Do not let a vendor's 10.0 drive an emergency response without checking reachability; do not dismiss a 5.3 that is internet-facing and pre-auth
- Legacy v2 scores still appear in feeds; map them coarsely and flag them rather than mixing them into v3/v4 statistics
EPSS Prioritization
EPSS (FIRST.org) estimates the probability that a vulnerability is exploited in the wild within 30 days: a score from 0 to 1 plus a percentile rank, derived from observed exploitation rather than expert judgment.
- EPSS is a triage input, not a severity replacement. High EPSS on an unreachable asset is still not urgent; low EPSS on an internet-facing pre-auth RCE is still urgent, because EPSS is a population average.
- Practical composite: CVSS ≥ 7.0 AND (EPSS ≥ 0.10 OR KEV). This captures most real-world exploited vulnerabilities while cutting queue volume sharply.
- Use the percentile in reporting ("top 5% of scored CVEs") and the raw score for thresholds.
- Refresh daily and re-rank the open backlog at least weekly.
- A brand-new CVE may have no score for hours — treat missing as unknown, not zero.
Composite Priority Model
priority = severity_weight(CVSS band)
+ exposure_weight(internet +3 / partner +2 / internal +1)
+ exploit_weight(KEV +4 / EPSS ≥ 0.5 +3 / EPSS ≥ 0.1 +2 / public PoC +1)
+ criticality_weight(Tier 1 +3 / Tier 2 +1)
- mitigating_control(compensating control in place -2)
Sort descending, apply SLAs by the resulting band, publish the formula so owners can see why their item ranks where it does, and re-tune weights quarterly against what actually got exploited in your environment.
CISA KEV is the curated list of CVEs with confirmed in-the-wild exploitation. Small, high-signal, mandatory. Any KEV entry present in your environment is an emergency. Ingest it as a daily feed diff, not a manual check.
Patch Management Lifecycle
Patching is a supply-chain operation. Automate the middle stages; never skip the last.
1 — Assessment. Identify affected assets by exact version, not product name. Confirm the vulnerability is reachable in your configuration (a disabled module is not exploitable). Read the advisory for prerequisites, regressions, and required intermediate patches. Note whether a reboot, restart, or schema change is needed. Record finding, asset, fix version, and window.
2 — Testing. Apply to a canary group (1–5% of fleet, one of each OS/role). Run smoke tests and critical-path suites. Watch error rates, latency, and crashes for 24–72 h. Check vendor known issues. Document the rollback procedure and artifact before touching production.
3 — Deployment. Wave the rollout: canary → 10% → 50% → 100% with automated health gates. Respect maintenance windows and freezes — a KEV critical overrides a freeze with incident-commander approval. Prefer immutable replacement over in-place patching in cloud; rollback becomes trivial. Where no patch exists, deploy a compensating control (segmentation, WAF rule, feature disable) and record it against the finding.
4 — Verification. Re-scan the specific asset and confirm the finding is gone. Confirm the running version, not just the installed one — containers and JVM apps often keep old artifacts. Close only with verification evidence attached. If verification fails, reopen and escalate: a falsely closed finding is worse than an open one.
Automation Patterns
- Dependabot / Renovate: automated update PRs; auto-merge patch/minor with green CI, require review for majors
- Ansible / Chef / Intune / Fleet: scheduled OS patch runs with reboot orchestration and canary inventory groups
- Container rebuild pipeline: nightly base-image refresh → rebuild → scan → deploy; the only sustainable way to keep images patched
- Idempotent remediation scripts: safe to re-run, log to stdout, exit non-zero on failure so the orchestrator can gate
- Guardrails: rate-limit concurrent reboots, require approval for large production waves, always provide a kill switch
Zero-Day Response Playbook
Triggered by a vulnerability with no vendor patch, credible reports of active exploitation, and a plausible path to your assets.
Phase 0 — Triage (first 2 h). Confirm the report is real (vendor acknowledgement, reputable research, KEV or CERT advisory) — premature and false disclosure are common. Identify affected products/versions and whether you run them. Assign one incident commander and one channel. Determine exposure: internet-facing count, authenticated-only count, whether the component is enabled.
Phase 1 — Contain (first 24 h). Apply the highest-fidelity mitigation available, in order: disable the vulnerable feature → network block/segmentation → WAF or IDS virtual patch → reduce service privilege. Increase logging and hunt for prior compromise (zero-days are often exploited before disclosure). If exploitation is confirmed, pivot to incident response: isolate, preserve evidence, follow the IR plan.
Phase 2 — Eradicate. Deploy the vendor patch as soon as it exists, ahead of normal change control, with an emergency change record. Without a patch, maintain the control and set a review date. Rotate any credentials or keys that could have been exposed. Verify with targeted scanning that the control holds.
Phase 3 — Recover and learn. Restore operations; remove temporary controls only after the patch is verified. Write the timeline (detection, decision, containment, patch). Add a permanent detection rule for the exploitation pattern. Update the asset inventory with anything found during the hunt. Feed lessons into scan policy, SLAs, and patch automation.
Communication. One authoritative channel; no written speculation. Executive updates every 24 h while active, even if unchanged. Never publish exploit details or internal exposure counts. Coordinate disclosure with the vendor.
CVE Monitoring and Alerting
| Source | Signal | Latency |
|---|---|---|
| NVD | Canonical record, CVSS, CWE, CPE | Hours–days after CVE |
| CVE.org / MITRE | CVE assignment and description | At publication |
| CISA KEV | Confirmed exploitation | Same day |
| FIRST EPSS | Exploitation probability | Daily |
| Vendor advisories (MSRC, Red Hat, Cisco, Oracle, Atlassian, GitHub) | Patch availability, versions | At release |
| GHSA | Open-source package vulns | Often before NVD |
| OSV.dev | Multi-ecosystem package vulns | Near-real-time |
| oss-security, full-disclosure | Early warning, sometimes pre-CVE | Earliest |
| Exploit-DB, Metasploit modules | Weaponization signal | Post-disclosure |
Pipeline: ingest on a schedule (daily minimum; hourly for KEV and vendor feeds) → normalize to a common schema (cve_id, products/CPEs, versions, CVSS vector, EPSS, KEV flag, published date, references, patch availability) → match against inventory by product + version range, not keyword → dedupe against existing findings (update, don't duplicate) → score with the composite model → notify the owning team with finding, fix, and deadline.
Alert design. Alert on match + severity, never on every published CVE. Suppress products you do not run — the primary source of alert fatigue. Group by asset owner so one team gets one digest. Include the action: patch version, workaround, or "not affected". Page on-call for critical/KEV; everything else goes to a ticket queue. Track alert-to-acknowledgement time.
Advisory discipline. Subscribe to vendor advisories for every product in inventory and maintain that list at onboarding. Review advisories even without a CVE — some vendors ship fixes silently. Watch end-of-life announcements: EOL software generates unfixable findings forever and needs migration, not patching.
Vulnerability Metrics
| Metric | Definition | Target signal |
|---|---|---|
| MTTR | Finding creation → verified closure | Down per severity band |
| Remediation rate | Closed / opened per period | ≥ 1.0 (backlog shrinking) |
| Exposure window | Median days critical/high open on internet-facing assets | Critical ≤ 3 days |
| SLA compliance | % closed within SLA | ≥ 90% critical, ≥ 80% high |
| Backlog aging | Open findings > 90 days by severity | Decreasing |
| Scan coverage | % of inventory scanned last cycle | ≥ 95% |
| Reopen rate | Closed findings that reopen | < 5% (higher = false closures) |
| False positive rate | Rejected findings / total | < 15% |
| Risk-weighted backlog | Sum of open priority scores | Decreasing |
| KEV exposure | KEV findings open, and their age | 0, always |
Cadence: daily new criticals/KEV, SLA breaches, scan failures; weekly queue by owner, aging outliers, expiring exceptions; monthly MTTR trends, SLA compliance, coverage, reopen rate, top recurring root causes; quarterly posture trend, maturity, investment asks.
Anti-patterns: raw finding counts reward noisy scanners; averaged severity hides the critical tail; MTTR alone can be gamed by early closure — pair with reopen rate; "scans run" is activity, not outcome; cross-team MTTR comparison without normalization is misleading.
SIEM / SOAR Integration
SIEM. Forward findings and asset context so exploitation alerts can be enriched with "this host has CVE-X open". Correlate EDR/IDS exploitation signals with open findings to prioritize live incidents. Build detections for exploit attempts against known-vulnerable services. Retain findings 1 year+ for audit.
SOAR. Automate ticket creation → owner routing → SLA timer. Auto-enrich with CVSS vector, EPSS, KEV, asset criticality, known exploits. Auto-close only on re-scan verification, never on a "done" comment. Auto-expire exceptions and reopen findings. Automate containment for confirmed exploitation (isolate host, disable account, block IOC) with approval gates on destructive actions. Keep playbooks idempotent so retries don't duplicate tickets.
Ticketing hygiene. One ticket per finding-asset pair, not per scan. Include CVE, asset, fix version, SLA deadline, verification command, rollback plan. Link to the change record for audit traceability. Use a dedicated queue/component so metrics are computable.
Scripts
scripts/cve_lookup.py— NVD lookup with CVSS vector, EPSS score, and KEV status; retries on NVD rate limits.scripts/prioritize.py— rank a JSON/CSV findings export with the composite model above and compute SLA due dates and overdue status.
Both are stdlib-only (Python 3.9+) and support --help. Pass an NVD API key via
NVD_API_KEY, never on the command line.
Pitfalls
- Scanner-first program: buying a tool and calling it a program. Without ownership, SLAs, and verification, findings rot in a dashboard.
- Severity sorting: a CVSS-sorted queue is dominated by unreachable library CVEs while an internet-facing medium with a public exploit waits at the bottom.
- Unaudited exclusions: "accepted risk" lists that grow forever. Every exception needs an expiry.
- False closures: closing because a patch was "deployed" without re-scanning. Containers and JVM apps frequently still run the old artifact.
- Credentialed scan blindness: a broken scan account drops coverage to near zero while dashboards stay green.
- Ignoring compensating controls: teams patch what is easy, not what is exposed. Segmentation and WAF rules are legitimate when documented.
- Alert fatigue: forwarding every new CVE to every team instead of matching against inventory first.
- No asset owner: unowned assets are where criticals live for years.
- EPSS as severity: it is a probability and a population average; it cannot see your architecture.
- Patching without rollback: an emergency patch that breaks production is an outage, not a security win.
- Activity metrics: scans run, tickets created, reports sent — none say whether exposure went down.
- Zero-day panic patching: fleet-wide deployment of an unverified patch within an hour, causing a second incident.
Verification
- Asset inventory has owner, criticality, and exposure for every asset
- Scan coverage ≥ 95% of inventory, with failed scans alerted on
- Credentialed scanning verified working on a sample host
- External scan covers all owned public ranges and domains
- SLAs defined per severity and exposure, with KEV override documented
- Prioritization uses CVSS + EPSS + KEV + asset criticality, formula published
- Every open critical/high has an owner and an SLA deadline
- Patch process includes canary, health gates, and rollback artifact
- Remediation verified by re-scan before ticket closure
- Exception register reviewed; all entries have expiry dates
- Zero-day playbook documented with a named incident commander role
- CVE/KEV/vendor feeds ingested on schedule with dedupe
- MTTR, SLA compliance, coverage, and reopen rate reported monthly
- SIEM enrichment and SOAR ticket automation tested end to end
- Audit evidence (scan reports, closure records, exceptions) retrievable