Enterprise Security in the Agentic AI Era

FIRST Vulnerability Forecast

3 min read

FIRST Vulnerability Forecast 2026

Source: FIRST — Vulnerability Forecast for 2026 (2026-02-11). FIRST is the Forum of Incident Response and Security Teams, the global body behind CVSS and the coordinated-disclosure norms most vulnerability programs follow.

Key Claim

CVE publication crosses 50,000 for the first time in 2026. The median forecast is roughly 59,000 disclosed vulnerabilities, and the model’s credible interval reaches well past 100,000. Patch, detection, and disclosure capacity must be planned against that volume rather than against last year’s count.

Methodology

FIRST built a statistical time-series model trained on post-2017 CVE data, chosen because disclosure patterns changed structurally after the CVE program’s mid-2010s expansion. The model reported 7.48% mean absolute percentage error on its 2025 yearly prediction, a calibrated track record rather than a one-off guess. The forecast projects the published-CVE count, not total latent vulnerabilities.

Notable Findings

  • Median 2026 forecast: 59,427 CVEs, with a 90% confidence interval of 30,012 to 117,673. FIRST calls 70,000-100,000 “an entirely realistic scenario.”
  • 2026 is the year the count crosses 50,000 for the first time.
  • Three-year outlook flattens after the 2026 spike: ~59,000 (2026), ~51,000 (2027), ~53,000 (2028). The model does not project runaway exponential growth.
  • The forecast does not model AI-driven discovery as a separate factor. It extrapolates from historical publication trends. AI-accelerated discovery, the force behind the Glasswing funnel and MOAK, is a plausible source of upside surprise the model does not explicitly price in.

Strengths and Weaknesses

Strengths. Authoritative, methodologically transparent, and calibrated against a measured prior-year error. FIRST has no product to sell against the number, which makes this a cleaner capacity-planning anchor than vendor forecasts.

Weaknesses. It forecasts published CVEs, which lag actual discovery and depend on CNA assignment throughput, a bottleneck the AI-discovery wave may itself break. If AI-discovered findings bypass or overwhelm the CVE pipeline, the published count could understate real exposure. The model is also blind to the AI-discovery factor by construction, so its flat 2027-2028 outlook should be read as a baseline, not a ceiling.

Relations

  • Supports VulnOps — quantifies the triage-volume problem the function is designed around. A team planning for ~59,000 CVEs (and possibly 100,000+) cannot route every finding through a human queue.
  • Complements Verizon DBIR 2026 — FIRST quantifies how many vulnerabilities arrive; the DBIR quantifies how fast the ones that matter get exploited. Together they bound the volume and the speed.
  • Supports CTEM — capacity-planning input for the Scoping and Prioritization stages.

AI-discovery upside is unpriced

FIRST’s model extrapolates historical publication trends and does not treat AI-accelerated discovery as a variable. If frontier-AI vulnerability discovery routes findings into the CVE pipeline at scale, the realized 2026 count could land near or above the 117,000 upper bound rather than the 59,000 median. The flat 2027-2028 outlook is a baseline that assumes the discovery process does not change shape — an assumption the rest of this wiki documents being broken.

Graph View

Backlinks