AI vulnerability prioritisation in 2026: EPSS, VPR, and Agent Val

CVSS-only prioritisation is dead because CVSS measures severity in an ideal attacker scenario, not in your environment. A CVSS 9.8 critical with no public exploit, no active exploitation, and only affecting an isolated dev server is lower priority than a CVSS 6.5 moderate that is actively exploited (in CISA KEV), has EPSS above 0.85, and affects an internet-facing payment processing system.

Teams adopting EPSS-weighted prioritisation with CISA KEV as an authoritative exploited-in-wild list report 60-80% reduction in weekly patch tickets. The 80% of tickets that disappear are CVEs that scored high on CVSS but have never been exploited in production. AI prioritisation layers asset context on top to produce the final actionable list. Unmanaged assets often carry the worst CVE posture without anyone realising it; see shadowitcost.com for the shadow IT exposure angle.

EPSS (Exploit Prediction Scoring System) is maintained by FIRST.org. Version 3 (current) predicts exploitation probability over the next 30 days using a model trained on NVD metadata, PoC publication signals (Exploit DB, GitHub), threat-intel feeds (scanning activity, underground chatter), and historical exploitation patterns.

Interpretation thresholds (from FIRST.org guidance, adapted to practical use):

EPSS scores shift as new threat data arrives. A CVE at 0.05 today can move to 0.80 when a PoC is published. Automate EPSS score refresh against your CVE inventory daily via the FIRST.org EPSS API (free, rate-limited).

Tenable VPR ships in Tenable One and Tenable Vulnerability Management. It combines CVSS with EPSS scores, asset exposure paths (is the vulnerable service reachable from the internet?), business-context weighting (production vs dev environment), and real-time threat-intel feeds from Tenable Research.

AI augmentation in 2026: natural-language remediation guidance (the dashboard explains what to patch and how in plain English, not just a CVSS score), exposure-path reasoning (showing the attack chain from the vulnerability to a critical asset), and business-context weighting that can incorporate asset criticality tags the security team maintains.

Pricing: Tenable One starts at approximately $30k-$60k per year for small to medium teams. Tenable Vulnerability Management (the standalone product, previously Tenable.io) has lower entry pricing for teams not needing the full Tenable One suite. Tenable is available on the Impact affiliate programme; pricing confirmed via Tenable's partner programme (April 2026).

Honest verdict: strong for existing Tenable customers and teams wanting a combined scan-plus-prioritisation platform. Cold-start friction is real if you are not already using Nessus or Tenable scanners. The VPR score is genuinely more actionable than CVSS alone, and the natural-language remediation guidance saves analyst time.

Qualys Agent Val is an agentic AI that validates exploitability at machine speed using TruLens threat context. The problem it solves: traditional vulnerability scanners detect all CVEs in the environment. Agent Val narrows that list to CVEs that are both (1) actually present and exploitable in your specific configuration and (2) actively weaponised based on current threat intelligence.

How it works: TruRisk score (Qualys' combined risk metric: CVSS, EPSS, CISA KEV, asset criticality) layers with TruLens (Qualys' threat-intel layer, incorporating exploitation signals from threat feeds) to produce a highly filtered remediation priority list. Agent Val then validates exploitability at machine speed against the asset's actual configuration, eliminating false positives from the scanner output before they reach the analyst.

Pricing: included in Qualys Total CyberRisk Management contract, which is custom-priced. Agent Val and TruLens are not available as standalone add-ons as of April 2026. Qualys' Total CyberRisk Management targets enterprise security teams with existing Qualys VMDR deployments.

Honest verdict: the most ambitious AI vulnerability-prioritisation product announced in Q1 2026. Early production in April 2026; enterprise deployments are limited. The approach of combining TruRisk scoring with agentic exploitability validation is sound. Watch for independent benchmarks on false-negative rates (CVEs Agent Val marks as low-priority that prove exploitable in practice) before committing a large contract.

Documented and working in April 2026 (source: Penligent April 2026 blog):

# Zero-cost AI vulnerability prioritisation

1. Scan: OpenVAS (free) or Nessus Essentials (free, 16 IPs)
   -> produces CVE list per asset

2. Enrich per CVE (free APIs):
   - NVD API: CVSS score, CPE, description
   - EPSS API (first.org): exploitation probability
   - CISA KEV JSON: known-exploited flag

3. LLM agent (Claude API or local Llama 4):
   Input: CVE data + asset context (criticality tag, network exposure)
   Output: prioritised remediation list with rationale

   Prompt pattern:
   "Given this CVE list with CVSS, EPSS, KEV status,
   and asset criticality, produce a prioritised
   remediation list. For each CVE, state:
   priority (P1/P2/P3), one-sentence rationale,
   recommended action."

4. Output: structured list, human team reviews P1s daily
   Cost: LLM API ~$20-$50/mo at typical SMB scan volumes

Limitation: no commercial threat-intel integration, no asset-graph exposure-path analysis, and no automated exploitability validation (unlike Qualys Agent Val). The OSS pattern works for teams with under 500 assets and no dedicated vulnerability engineer. Above that scale, commercial tooling's automation pays for itself. See open-source tools for the broader OSS security stack.