The speed it promises already outruns most defenders: public exploit code has trailed the CVE by under a fortnight, at the median, in every complete year since 2015, while most estates take weeks to remediate. The discovery it promises exists now: machine-found vulnerabilities in real codebases are a present-tense capability, and on most estates that dial outweighs the other three together. The rest, the share of bugs weaponised and the tempo past the foothold, runs on dials rather than baked-in assumptions. Set your estate against all four and the Monte Carlo simulation reruns on every change, down to one number: the probability that you are compromised within twelve months.
Eight access classes are counted: …. Three of them need a published vulnerability. The rest do not, and no change window closes them. Not counted: …. Vulnerability exploitation accounts for 20% of breaches in the DBIR: a minority of routes, and not the largest. Every figure here is a lower bound.
Conditions you face rather than choices you make. All are editable here.
Place yourself on each axis, then adjust any parameter. Every figure is recalculated as you go.
The routes that need no vulnerability. Four of the eight access classes scale with how many people can authenticate, not with how many systems you run.
Day zero is CVE publication. Everything to the left of it is exploit code that precedes the record — mostly publication lag, not zero-days. The hatched region is remediation that completes with exploit code already public; the badge is that probability across all remediation, drawn window or not.
The horizontal axis is days from CVE publication. Day 0 is the vertical rule, and the shaded band to its left is time when no CVE record existed yet. Most of that band is not a zero-day. It is publication lag: exploit code is already public and mass scanning has already followed it, so the model treats that time as fully exposed rather than discounting it the way it discounts a genuine pre-disclosure exploit. The same series puts 98.5% of exploits ahead of publication in 2000, at a median of 44 days. That measures how slowly records were published, not what attackers could do.
Two distributions mirror off the centre line. Above it, remediation complete is when your remediation lands, day by day. Below it, exploit available is when public exploit code appears, drawn twice: as a faint daily density, and as the bold cumulative curve. The cumulative curve reads against its own scale, which runs from 0% at the centre line down to 100% at the bottom.
Unremediated exposure is the hatched region. For each day it is the part of that day's remediation for which exploit code already existed, so it is the share of the remediation curve that lost the race.
The two large figures answer different questions. The marker on the cumulative curve at day 0 is the share of exploits that are already public when the CVE record lands. The badge is the share that arrive before your remediation completes, which is always the later test of the two and therefore always the larger number. The percentages at the edges are exploit or remediation mass falling outside the drawn window, labelled there rather than folded into the end bins, which would invent a spike that is not in the data.
Each stage is a subset of the one above it. The stream entering the top is every severity band, not just Critical: you set the Critical count, and the High, Medium and Low volumes beside it are derived from published ratios, each with its own rate of acquiring exploit code and of being exploited. Counting Criticals alone would discard two thirds of everything known to be exploited. The two largest reductions happen outside the remediation process: most published vulnerabilities land on a version or configuration that is not exposed to them, and most of the rest never get a working exploit.
More than half of first compromises on a typical estate involve no CVE, and none of them races a change window. Each is an arrival stream against a set of controls: filtering thins what reaches a person, authentication decides what a click or a stolen credential is worth, privilege decides what a valid account reaches, and configuration assurance closes what was never vulnerable, only reachable. Where the first two bars are equal, nothing this estate does thins arrival — the control acts only after the adversary is at the gate. Every rate here is judgement, not measurement; what anchors them is the mix they produce, held to the reported breach population drawn in chapter 06.
The largest single control on this half of the model is authentication, and it is a ladder rather than a dial — an adversary uses the weakest path in, not the best one available.
Each bar varies a single parameter across its plausible range, holding everything else fixed. Read the ranking rather than the ratio between bars: bar length depends on the range chosen for each parameter. The baseline marked on the chart is recomputed at this chart's own trial count with the uncertain inputs held central, so it can sit a point or two from the headline.
A compromise is an adversary establishing a foothold. It becomes an incident when detection and response fail to contain it before the adversary reaches its objective. The two respond to different parameters, and that difference is what should drive priority.
Two bars need a caveat. Detection is flat against compromise because detection only enters the model once a foothold exists, so controls that block exploitation outright sit outside the model rather than measuring as zero. Headcount is a measure of size rather than a control, and the routes that scale with people scale less than proportionally, because one lure reaches every mailbox at once.
The five changes to your configuration with the largest effect on annual probability of compromise, at your current settings.
Class of the first compromise, counted once per simulated year. Only opportunistic exploitation turns fully on your change window; a targeted campaign turns partly on it, and the other six not at all. That is why the ranked actions above are not simply patch faster. The five classes that need no vulnerability rest on judgement rather than measurement, so the pair of bars beneath shows the only check available on them: this estate's mix of routes against the reported breach population it is calibrated to.
Proportion of simulated years without compromise, by day.
BOD 26-04 names adversary use of AI as a reason the window between patch release and exploitation may narrow. That is four separate claims with different evidence behind them and different consequences, so each gets its own dial. All four default to the measured record.
is the one the phrase usually means, and of the three that can move compromise at all it moves it least. Not because the clock has nowhere left to go: the measured median has stayed under a fortnight in every complete year since 2015, and it has lengthened rather than shortened, on a sample small enough that the direction is as much as it can show. Because at that speed exploit code is already ahead of most remediation, so a faster clock mostly shortens races the exploit was winning. is worth more than twice as much on the default estate, because breadth adds races where speed only shortens them.
is flat against compromise because speed from foothold to objective cannot change whether you were breached, only whether anyone reached it in time. Switch the readout to incidents and it lifts off the floor. Even there the curve understates where it bites: against a 24/7 SOC, full tempo removes roughly half of the margin that detection buys over having none. A faster adversary does not beat detection on any single intrusion. It devalues the investment.
The steepest curve here is the last one, and it is not a clock at all. asks how many vulnerabilities are found in software you already run: the size of the stream rather than the speed of anything downstream of it. Machine-found bugs in real codebases are a capability that exists today rather than a forecast, which makes this the least speculative of the four and, on most estates, the largest.
The conventional approach counts Critical-rated CVEs. The published record does not support it. Below is the full scored CVE population, bucketed by CVSS band, against the share of each band that has reached the confirmed-exploited catalogue.
| CVSS band | Scored CVEs | Confirmed exploited | Rate |
|---|
This position is now codified. CISA Binding Operational Directive 26-04 (10 June 2026) supersedes BOD 19-02 and BOD 22-01 and removes severity as the trigger entirely. It prioritises on four SSVC decision points: presence in the KEV catalogue, public exposure of the asset, whether exploitation is automatable, and the technical impact obtained. Remediation tiers run from three days plus forensic triage down to “fix on system upgrade”. The published heuristic for automatable is a public proof-of-concept achieving remote code execution that runs reliably, which is the same test this model applies.
The Critical-rated population is growing faster than the CVE population that contains it. A model anchored to criticals per year therefore drifts more pessimistic each year for reasons unrelated to the estate being assessed. This is the second reason severity is not used as an input here.
Part of that increase reflects scoring policy rather than vulnerability reality. CNAs self-assign scores, and the share each rates 9.0 or above ranges from near zero to roughly two in five against the same scale. NVD has also stopped analysing a growing share of published CVEs.
Every input carries one of three labels, kept separate rather than blended. measured is read directly from the vulnerability data shipped with this page and can be reproduced offline. reported is a dated published figure: cited, but dependent on someone else's method. assumed is judgement, used where no public measurement exists. The last two are redrawn from their stated range throughout every run, which is where the uncertainty bands come from. If you disagree with a number, the range beside it is the thing to argue with.
Both exports report your estate rather than this page's argument: what you configured, the resulting probability with its uncertainty band, the funnel and route split behind it, what moves the number most, and the ranked actions with the reduction each buys here. The last slide carries a link to this exact configuration, so a colleague opens the run rather than rebuilding it. Every slide repeats the scope limit in its footer: this is a lower bound, never a risk assessment for a named organisation.
Portrait slides at 1080 × 1350, kept short enough to read on a phone. Upload it as a document post and each page becomes a swipeable slide.
16:9 at projection size, and two slides longer: the parameter values your selections produce, and the method. Text is selectable, so a figure can be quoted without being retyped.
Built in the browser from the run currently on screen. Nothing is uploaded. The charts are the ones drawn above, so scroll through the page first if you want every chart in the deck.