Vulnerability Management Solved Finding. It Never Solved Deciding.
Scanners find everything. Scores rank everything. Then every finding still waits for a human to decide what to do about it, one ticket at a time. Where vulnerability management breaks at 2026 volumes, and what replaces the queue.
TL;DR
- Vulnerability management is the practice of finding, ranking, fixing and verifying security weaknesses across an organizationβs systems. The standard five-stage definition never says who decides, or how often the decision gets remade.
- The model breaks at volume. 48,185 CVEs were published in 2025, a record. 2026 had produced 65,786 by September 15, within a few hundred of FIRSTβs forecast for the whole year.
- Scanners solved discovery and scores solved ranking. The backlog persists because every finding still needs a fresh human decision and an ownership handoff. Better prioritization shortens the list without reducing the number of decisions.
- Exposure management widens the job past CVEs to misconfigurations, identities and attack paths, and carries findings through to a verified fix. Most platforms using the term stop at a better assessment.
- Decision-Based Exposure Management changes the unit of work from the task to the decision: a human decides once per class of exposure, and agents execute that decision continuously across every current and future instance, under guardrails the human set.
Mondayβs scan finished at 6:40 a.m. and delivered 1,900 new findings. About 400 are the same glibc CVE on the same container base image, owned by three different teams depending on which CMDB field you trust. Each one becomes a ticket. Someone will open ticket 401 with fresh eyes, as if the first 400 never happened.
What is vulnerability management?
Vulnerability management is the continuous practice of identifying security weaknesses across an organizationβs systems, ranking them by risk, fixing or mitigating them, and confirming that the fix held. The textbook vulnerability management lifecycle runs in five stages: discover assets and scan them, assess and classify what the scan found, prioritize by severity and exploitability, remediate or apply a compensating control, and verify closure. A reporting loop wraps around it so the program can show progress.
That definition is accurate and incomplete. It describes five stages as steps and names no decision-maker for any of them. In practice the decision is made per finding, by whoever picks up the ticket, on the day they pick it up. Two analysts can look at the same CVE on two identical servers and reach different conclusions, and the process has no way of noticing.
Roughly 6% of all published CVEs have ever been observed with exploitation activity in the wild, according to a 2024 Cyentia Institute and FIRST study. The job was never to fix everything. It was to decide, for each finding, whether it belongs to the 6% or the 94%, and what to do about it. The standard definition is silent on how that decision gets made and who owns it, and Onitβs earlier piece on the 30-year-old ownership problem in exposure management covers what that silence costs.
Why vulnerability management breaks at scale
It breaks because the unit of work is the individual finding, and findings grow faster than the people who decide about them. This is not one teamβs staffing problem. The volume curve is structural and it is steepening.
48,185 CVEs were published in 2025, a record, according to CVE Program data. 2026 has already passed it with a quarter of the year left: 65,786 CVEs were published between January 1 and September 15, 2026, within a few hundred of the roughly 66,000 that FIRST forecast for the entire year in its June update. The 2026 count to date is about 65% higher than all of 2024, which closed at 39,962.
The must-fix-now subset grows too. CISAβs Known Exploited Vulnerabilities catalog stood at 1,710 entries on September 14, 2026, after 245 additions in 2025 and 226 so far in 2026. A team that ignores everything except KEV is still absorbing a new fix-now decision roughly every working day.
A program built on one human judgment per finding needs headcount to scale with that curve, and no security team gets it. So the queue grows, SLAs slip, and the same judgment call gets re-made thousands of times a year by different people with different context. It is a decision-throughput problem wearing the costume of a volume problem.
Scanning and scoring are solved. Deciding is not.
Credit where it belongs. The scanners this audience already runs, Tenable, Qualys, Rapid7, find things reliably across hybrid estates that were hard to even inventory a decade ago. CVSS gave the industry a shared language for severity. EPSS and CISA KEV added exploitation likelihood and observed exploitation on top of it. Ticketing systems track every finding to closure with an audit trail. Not knowing what you had used to be the problem. That problem is gone.
What none of it changed is what happens after a finding lands. Every one still requires a fresh human decision (is this exposure acceptable here, and if not, what do we do about it) and then an ownership handoff to someone with the authority to act. Those two steps are where the time goes.
From Onitβs own work with enterprise programs, roughly half of remediation tickets bounce back because they reached the wrong owner, and about 30% stall because the engineer who received them did not have enough context to act. Those are Onit field observations and most vulnerability management leads will recognize them from their own queue.
The failure mode is not hypothetical. Onit was founded after state-sponsored attackers got into co-founder Ofer Amitaiβs previous company through a single known vulnerability. Not a zero-day. A flaw that had been found and scored, then left in a backlog no one owned, until someone else found it first.
CVSS and EPSS are useful signals, not a strategy
CVSS scores the vulnerability. EPSS scores the world. Neither one scores your environment.
CVSS describes how bad exploitation would be if it happened: attack vector, complexity, privileges required, impact. It says nothing about whether it can happen on your network. EPSS, published daily by FIRST, is a 0 to 1 estimate of the probability that exploitation activity against a CVE will be observed in the wild over the next 30 days, based on attempted exploitation seen across FIRSTβs data partners. It is not confirmed compromise, and it is not about your instance. FIRSTβs own documentation says EPSS βdoes not account for any context for specific instances, any compensating controls, or any factor specific to an organizationβs environment.β The current model is EPSS v5, publishing since June 15, 2026.
Both signals are real improvements in ordering the list. That is what risk-based vulnerability management does: it layers exploitation likelihood and asset criticality onto raw severity so the top of the queue is more likely to matter. Onitβs post on why perfect prioritization still doesnβt shrink the backlog covers where that stops.
Where it stops is here: a better score narrows the list, and a narrower list is still decided one item at a time. Filtering hides the backlog. It does not retire any of it. Ten thousand findings ranked perfectly are ten thousand decisions. Ranked badly, they are also ten thousand decisions.
How long does vulnerability remediation actually take?
Weeks to months, and the fast lane is getting slower. The figures below measure different populations and must not be blended into one βaverage.β
| Figure | What it measures | Source |
|---|---|---|
| 54.81 days | Mean time to remediate, high and critical severity application and API vulnerabilities, 2025 data | Edgescan Vulnerability Statistics Report 2026 (11th edition) |
| 53 to 123 days | Mean time to remediate by industry, from software (fastest) to construction (slowest) | Edgescan 2026 |
| 43 days, up from 32 | Median time to full remediation of CISA KEV vulnerabilities, across Verizonβs scanner-partner dataset | Verizon 2026 Data Breach Investigations Report |
| 26%, down from 38% | Share of KEV vulnerabilities fully remediated in 2025 | Verizon 2026 DBIR |
| 243 days | Remediation half-life across all flaws and scan types (time for half of discovered flaws to be closed) | Veracode State of Software Security 2026 |
The Edgescan figure is the cleanest single headline: for high and critical application and API vulnerabilities, the average mean time to remediate in 2025 was 54.81 days. The DBIR figure is the more alarming one for what it tracks. That 43-day median covers the subset every program already treats as urgent, and it moved the wrong way by almost two weeks in a year.
Set those against the attackerβs clock. Mandiantβs M-Trends 2026 puts the mean time to exploit for 2025 at an estimated negative seven days, meaning exploitation, on average, now begins before a patch exists. The exploit arrives a week early. The fix arrives six weeks late. That gap does not close by asking the same team to move faster through the same queue, and Onitβs MTTR pillar covers why MTTR remediation resists improvement in detail.
From vulnerability management to exposure management
Vulnerability management asks a narrow question: which CVEs do we have, and how bad are they. Exposure management asks the real one: what in our environment is actually exposed, and can an attacker reach something that matters. It widens the aperture past CVEs to misconfigurations, identities and attack paths, and it carries a finding through to a verified fix rather than stopping at a ranked list. This is the shift Gartner formalized as Continuous Threat Exposure Management in 2022, and it is where the category is already moving. The full definition lives in Onitβs exposure management pillar; this page only needs the boundary.
The major scanners and cloud security vendors are already saying βexposure, not vulnerabilities.β The term itself is no longer a differentiator. The first thing most of them sell under it is consolidation: every scanner normalized into one deduplicated view, which is a better dashboard, and Onit has written before about why unifying tools isnβt the same as unifying the fix.
What still separates platforms is what happens after the assessment. Most stop there. They widen what you can see without changing what happens after you see it. The backlog, the ownership guessing and the manual resolution are all still in place, now with richer findings feeding them.
What Decision-Based Exposure Management changes
Decision-Based Exposure Management changes the unit of work from the task to the decision. A human makes one judgment about a class of exposure, and agents execute that judgment continuously across every instance that matches, today and as new ones appear. Humans define judgment. Agents execute.
Take the 400 glibc findings from Mondayβs scan. In a task-based program they are 400 tickets. Under Decision-Based Exposure Management they are one class, and the decision might read: any KEV-listed CVE on an internet-facing asset is patched within 24 hours, with a compensating control applied immediately if the patch window cannot be met. That covers the 400 instances today and the 60 that arrive in next quarterβs base image refresh, with nobody re-triaging them. Decide once. Resolve forever.
What makes a class defensible is the context scores cannot supply, checked in a fixed order. Is the instance reachable: does an attacker have a plausible path to it, and does the vulnerable code actually run. Is it exploitable here, or has a compensating control already closed it. What does it reach that the business cares about: the data it processes, what it connects to, what breaks if it stops, which regulation covers it. And who has the authority to fix it. Those four answers produce a defensible short list, and a short list is still a backlog. The decision is what retires it.
The model has five working parts:
- The unit of work is the decision, not the task. One human judgment per class of exposure replaces per-instance triage.
- One decision resolves thousands of exposures, present and future. New matching instances are covered the moment they appear.
- The decision persists as an operating rule, tightened or loosened as the environment changes. Resolve. Donβt re-solve.
- Ownership resolves from live context across the CMDB, identity systems, ticketing and collaboration tools, rather than a static mapping that was stale the day it was entered.
- What accumulates is institutional knowledge. Every decision keeps working after the analyst who made it has moved on.
Because this puts agents in front of production systems, four questions get direct answers. What the agents do: find every instance matching a decision, resolve the current owner, prepare the fix or control, and carry it through. What data they use: scanner output plus CMDB, identity, ticketing and collaboration data from the organizationβs own stack, and customer data never trains models. Where human review sits: humans write the decision and its guardrails, the system proposes and humans approve, every action is logged and reverses in one step. Who authorizes disruptive actions: the human who set the policy, by deciding which actions run under standing approval and which need a sign-off each time. Onit holds SOC 2 Type II, ISO 27001 and ISO 42001 for AI management systems.
This is the operating model Onit runs, working with Fortune 500 companies including Honeywell and Marvell. The largest deployment runs more than 50 million active exposures across roughly 700,000 assets with a team of about 20, which is the scale at which one decision per finding stops being an option.
[PROPOSED QUOTE, for Elad Ben Meir, CEO and co-founder, to approve or rewrite:] βYou cannot triage at human speed against an attacker who exploits at machine speed. The only thing that scales is a decision that has already been made.β
Keep reading: Onitβs post on what autonomous remediation really means walks through where the trust boundary sits when an agent carries out a standing decision.
Where does Decision-Based Exposure Management fit with CTEM?
Gartner introduced Continuous Threat Exposure Management in 2022 as a five-stage program: scoping, discovery, prioritization, validation and mobilization. It is the analyst-endorsed operating model for exposure management, and Decision-Based Exposure Management runs inside it rather than alongside it.
The first four stages are largely served by tooling most enterprises already own. Scanners and attack surface tools cover discovery. Risk-based scoring covers prioritization. Breach and attack simulation and pentest platforms cover validation. Mobilization, the stage where a validated exposure is actually driven to a fix by the people who can make it, is where programs stall, because it is the stage that depends on ownership, context and a decision.
Decision-Based Exposure Management is a mobilization layer. It takes the output of the first four stages and converts it into standing decisions that agents execute. Onitβs post on where CTEM programs actually stall covers each stage in depth.
Vulnerability management tools versus exposure management platforms: what to look for
A vulnerability management tool finds and scores. An exposure management platform sits on top of that output and orchestrates it into continuous execution. The two are not substitutes, and a platform that asks you to rip out a scanner you already trust is solving a problem you do not have.
| Vulnerability management tool (Tenable, Qualys, Rapid7) | Exposure management platform (Onit) | |
|---|---|---|
| What it ingests | Your asset inventory and network | Scanner output plus CMDB, identity, ticketing and collaboration data |
| What it produces | Findings with severity scores | Standing decisions, executed |
| Unit of work | A finding | A class of exposure |
| Who acts on a new finding | An analyst, per finding | An agent, executing a human-approved decision |
| What happens to the next identical finding | Re-triaged from scratch | Already covered by the standing decision |
| Relationship to your scanner | Is the scanner | Augments the scanners you already run |
| What it measures | Findings open and closed | Decisions in force, classes retired, time to verified fix |
Onit integrates with more than 100 tools including Tenable, Qualys, CrowdStrike, ServiceNow, Jira and Slack, over APIs, with no endpoint agents to deploy. Your scanner already finds these. The platform is what happens after.
The first steps toward this model do not require a platform, and they hold up as vulnerability management best practices whether or not you ever buy one:
- Pull the last 90 days of remediation tickets and group them by exposure pattern rather than by CVE. Three to five classes will account for most of the volume.
- For each class, write the standing decision in one sentence: the condition, the action, the time bound.
- Define the guardrails: which actions run under standing approval, which need a sign-off each time, and what triggers a rollback.
- Hand execution of those classes to automation, or to a platform built for it, and keep humans on the exceptions.
- Change what the program reports. Decisions in force, classes retired and time to verified fix tell a board about risk reduction. Tickets closed tells it about activity.
For a CISO, the case is not that this is bold. It is that a standing decision with an audit trail and a rollback is easier to defend than ten thousand individual judgments nobody can reconstruct after the fact.
The queue was never the job
Vulnerability management was built to empty a queue, and the queue is a symptom. It is what you get when the same decision is made ten thousand times by ten thousand different people on ten thousand different days. Ticket 401 should never have existed. The decision that closes it was available at ticket one.
Security by decision. Book a demo of Onit to see a standing decision written and executed, or read more on Onitβs approach to agentic exposure management.
Next reads
- Exposure Management Was Supposed to Fix a 30-Year-Old Problem. It Hasnβt.
- Why MTTR Stays High
- Risk-Based Vulnerability Prioritization: Beyond CVSS
- CTEM Framework to Execution
Frequently asked questions
A vulnerability management program combines the tooling, the process and the ownership model: scanners for discovery, a scoring method for prioritization (CVSS, EPSS, CISA KEV, asset criticality), defined remediation SLAs by severity, a ticketing workflow to track fixes to closure, and reporting to show progress. The part most programs leave undefined is who decides what to do about each class of finding, which is why two programs with identical tooling can perform very differently.
The standard vulnerability management lifecycle has five stages: discover (inventory assets and scan them), assess (classify what was found), prioritize (rank by severity, exploitability and asset criticality), remediate (patch, reconfigure or apply a compensating control), and verify (confirm the fix held). Most programs add reporting as a sixth, and mature programs run the cycle continuously rather than on a quarterly scan schedule. The vulnerability management process describes steps; it does not say who owns the decision at each one.
Vulnerability management finds and ranks CVEs on assets. Exposure management widens the scope to misconfigurations, identities and attack paths, adds business context, and carries findings through to a verified fix. Gartnerβs Continuous Threat Exposure Management framework is the analyst model for the wider practice. Decision-Based Exposure Management is Onitβs operating model for executing it: one human decision per class of exposure, executed continuously by agents.
Risk-based vulnerability management prioritizes findings by likelihood of exploitation and business impact rather than by CVSS severity alone, typically using EPSS, CISA KEV and asset criticality. It produces a better-ordered queue. It does not reduce the number of findings a team has to decide about, because each item on the shorter list is still handled individually.
Decision-Based Exposure Management is an operating model in which a human makes one decision about a class of exposure, and agents execute that decision continuously across every current and future instance that matches. Decisions persist as operating rules, ownership is resolved from live organizational context, and every agent action is logged, approval-gated and reversible in one step. It is the model Onit Security runs.
No. Scanners from Tenable, Qualys, Rapid7 and others handle discovery and scoring. An exposure management platform ingests their output, adds ownership and business context, and executes decisions on top of it. Onit integrates with existing scanners over APIs and does not deploy endpoint agents.