# EPSS vs AI exploit prediction: 14-day and 3-day patch clocks in 2026

Robert Chen · November 18, 2026

> EPSS ≥ 0.5 sets a 14-day patch SLA; KEV plus internet-facing exposure cuts it to 3 days. How the 2026 patch clocks work and where AI exploit prediction fits.

| Takeaway | Detail |
| --- | --- |
| EPSS ≥ 0.5 triggers a 14-day patch SLA as the defensible default for externally auditable change-control workpapers. | EPSS is the only daily-updated, externally auditable exploit-likelihood signal; threshold set at 0.5 or your calibrated cutoff. |
| KEV-listed, internet-facing vulnerabilities get a 3-day fast-lane patch window. | Condition: CISA KEV membership plus internet-facing exposure; overrides the 14-day SLA. |
| Vulnerabilities outside both KEV and EPSS ≥ 0.5 exit the SLA entirely until a signal change occurs. | No patch clock starts; re-entry requires a new KEV listing or EPSS crossing the calibrated threshold. |
| AI exploit prediction tools are noted but lack the auditability required for workpaper citation. | Grounding sources describe AI assistive roles; none provide the daily, auditable feed EPSS offers. |

This guide sets the patch clock by evidence class: KEV plus internet-facing exposure gets 3 days, EPSS ≥ 0.5 gets 14 days, and all other vulnerabilities exit the SLA until a signal changes.

The condition that changes the answer is whether the vulnerability is KEV-listed or reaches the EPSS threshold, because only those signals are auditable and daily-updated.

![EPSS vs AI exploit prediction](https://static.mm-ais.com/article-images-ai/epss-vs-ai-exploit-prediction-14-day-and-ai-fed01e2c.jpg)

## How EPSS and AI scores actually differ

The FIRST.org Exploit Prediction Scoring System publishes a daily-updated probability score from 0 to 1 for every CVE, indicating the likelihood the vulnerability will be exploited in the wild within the next 30 days. Because the model is versioned and the data source is public, you can snapshot the exact score on the day you make a patch decision and reproduce that number verbatim in an audit workpaper later. This transparency is what makes EPSS defensible in a change-control context; no opaque proprietary algorithm is standing between you and your evidence.

Vendor AI-driven exploit-prediction features, such as Tenable’s Predictive Prioritization or IBM Concert Protect’s proof-of-exploitability capability, layer machine learning on top of proprietary telemetry to rank findings, but the model inputs, update cadence, and underlying logic remain opaque and unaudited. When you rely on these scores, you are trusting a black box whose training data and version history are not published, meaning you cannot reproduce the exact number that justified a 3-day or 14-day patch clock during a later review. This mechanistic difference—published probability versus opaque likelihood—is the single factor that determines whether a signal can survive audit scrutiny.

To set your patch clock by evidence class, start with the KEV list: any CISA-known-exploited vulnerability on an internet-facing asset enters a 3-day remediation window with no exceptions. If a vulnerability is not KEV-listed but carries an EPSS score at or above your calibrated threshold—commonly 0.5—assign a 14-day SLA. Any finding that falls below the EPSS cutoff and is not KEV-listed exits the SLA entirely and must wait for a future signal change before re-entering the prioritization queue. This three-tiered clock ensures that only the highest-certainty evidence drives accelerated remediation.

Before you commit to any threshold, pull and date-stamp five figures from your environment: the current EPSS score for the CVE, the KEV status (present or absent), the asset’s internet-facing flag, the vendor AI score if one exists, and the model version of the EPSS data you are referencing. Record each value with a timestamp so that, come audit time, you can trace exactly which evidence class triggered which clock and recompute whether the decision remains defensible. If any of these figures are missing or the EPSS model version cannot be verified, the finding defaults out of the SLA until the signal is resolvable.
![How EPSS and AI scores actually differ — EPSS vs AI exploit prediction](https://static.mm-ais.com/article-images-ai/epss-vs-ai-exploit-prediction-14-day-and-ai-a4235878.jpg)

## The evidence: what each signal proves

When you are asked to defend a patch window in a change-control workpaper, the strongest position is to anchor every decision to a signal whose provenance you can point to and whose output you can recompute. FIRST’s Exploit Prediction Scoring System (EPSS) is currently the only externally auditable, daily-updated exploit-likelihood signal that meets that bar. The empirical case is published on first.org/epss: the model’s performance data show that exploitation likelihood outperforms the CVSS base score alone as a prioritization signal. When you cite that workpaper, you are not quoting a vendor benchmark; you are pointing to the published research that underpins the score.

Benchmark Authority frames exploitability signals precisely, and its language is worth reusing when a reviewer challenges why a high-EPSS finding receives a 14-day window instead of an immediate 3-day one. The guidance states that exploitability is “an input to prioritization, not a verdict,” and that such signals “represent an assessment of likelihood and accessibility, not proof of compromise.” That distinction is the exact defense for treating EPSS ≥ 0.5 (or your calibrated cutoff) as a 14-day trigger while reserving the 3-day fast lane for confirmed compromise evidence.

The convergence of evidence is therefore hierarchical. Exploitation evidence (CISA KEV membership) outranks prediction (EPSS probability), which in turn outranks inference (vendor AI scores). Each source carries its own framing: KEV is proof of in-the-wild exploitation, EPSS is a versioned daily probability from FIRST’s model, and AI scores are opaque, unaudited likelihood outputs. When you write the workpaper, cite each source by the framing it uses so the reviewer can see why one signal earns a clock and another does not.

As detailed in the checkpoint above, date-stamp the EPSS score, model version, KEV status, and internet-facing flag at the moment you open the change request. That checklist ensures the SLA you assign is traceable to the evidence class you identified, not to an arbitrary calendar entry.

What the evidence does not establish is the edge case where a high-EPSS, non-KEV vulnerability sits on a segmented internal host. In that scenario, neither the 3-day nor the 14-day clock applies automatically; the finding exits the SLA entirely and re-enters only when a signal changes. Treating that exclusion as a rule, rather than an exception, keeps the workpaper defensible when an auditor asks why a specific CVE was deferred.

## 14-day vs 3-day: options compared

The clock you attach to a vulnerability should follow the evidence class behind it, not the raw severity number a scanner prints. Two lanes are worth defending in a change-control workpaper — a 3-day emergency lane and a 14-day standard lane — and each has exactly one signal that earns it. Anything that qualifies for neither lane should drop out of the SLA rather than inherit a default window by inertia.

The 3-day lane has a single winner: a CISA KEV-listed CVE on an internet-facing asset. This is the only combination that pairs confirmed, in-the-wild exploitation with genuine external exposure, and that pairing is what justifies emergency change velocity. A KEV entry on an internal-only host does not clear this bar — exposure is half the test. When both conditions hold, you are not forecasting risk, you are responding to demonstrated risk against a reachable target.

The 14-day lane goes to any asset whose EPSS score meets or exceeds your calibrated cutoff. FIRST publishes EPSS as a daily probability, and the working range most teams land on sits between 0.1 and 0.5, sliding with asset criticality — crown-jewel systems warrant the lower end, more tolerant ones the higher. Treat that range as a starting hypothesis, not a setting: pull FIRST's published performance curves and pick the threshold that matches your tolerance for missed exploitation, rather than inheriting whatever default a vendor ships.

The table below is the whole policy in one view. Note the last row: vulnerabilities matching none of these signals do not get a longer window by default — they exit the SLA and re-enter only when a qualifying signal changes.

| Signal class | Clock | Why it wins |
| --- | --- | --- |
| KEV-listed CVE on an internet-facing asset | 3 days | Confirmed exploitation meets real external exposure |
| EPSS ≥ your calibrated cutoff (0.1–0.5) on any asset | 14 days | Auditable, daily-updated probability clears the bar |
| AI-only priority score (VPR-style, Concert Protect) | No clock of its own | Reorders the queue; not independently auditable |
| None of the above | Exits the SLA | Re-enters only on a qualifying signal change |

## Numbers that matter in 2026

Before setting any patch clock, pull the EPSS score and percentile for your top-100 backlog CVEs from the FIRST EPSS API on the day you finalize policy, and record the model version returned in the response header. Scores shift daily; a threshold derived from last week’s feed misroutes tickets when the underlying probability moves. Treat the snapshot as a dated artifact and attach it to the change-control workpaper so any later reviewer can recompute the same values.

Verify which EPSS model version your feed serves before comparing against any published cutoff. Check the version string returned by the FIRST EPSS API on the day you pull scores and note the date of retrieval; if your tooling surfaces an older model version than the one FIRST currently publishes, the numeric threshold you adopt may not be comparable to guidance written for the current release. Confirm the version, record it in the workpaper, and only then map your operational cutoff to the published band.

Count how many of your open criticals appear in CISA’s Known Exploited Vulnerabilities list. This single count drives the 3-day fast lane: every KEV-listed asset that is internet-facing moves immediately, regardless of EPSS score. Record the KEV count alongside the EPSS snapshot;

## What the evidence does not establish

The evidence class you rely on determines whether a ticket earns a clock at all. KEV membership on an internet-facing asset grants the 3-day patch clock; a KEV listing on an internal-only asset takes the 14-day lane, as in the worked example above — but verify the KEV entry actually covers the product and configuration you run before assigning either clock. EPSS provides a daily-updated likelihood probability that can shift sharply after new exploit activity is reported; a 14-day SLA set at a given threshold remains defensible only if the score is re-evaluated each day, with upward movement promoting the ticket forward. AI-powered exploitability scores operate as opaque, unaudited outputs from vendor-specific models and carry no independent clock of their own — they cannot displace the EPSS or KEV clocks without an accompanying signal change. The rule set holds: KEV plus internet-facing triggers the 3-day fast lane; EPSS at or above your calibrated cutoff triggers the 14-day lane; any classification outside these two buckets re-enters the SLA only when a new signal arrives.

Where the two-clock rule breaks down is defined by edge cases in product scope, score volatility, and model provenance. A KEV listing for a product you do not run in the affected configuration forfeits the 3-day acceleration, regardless of CVE membership alone; the KEV product scope field must be inspected, not assumed. EPSS scores are prone to sharp jumps when a new exploit is publicly published, meaning a threshold set at score X at the start of a cycle can become invalid before the 14 days lapse — daily re-evaluation is the safeguard, and any upward movement promotes the in-flight ticket forward rather than demoting it mid-cycle. AI models may surface exploitability assessments, but without the versioned, daily probability structure of EPSS or the exploitation evidence of KEV, they offer no standalone SLA trigger and must be treated as supplementary context rather than a clock-setting signal.

As established in the evidence hierarchy above, KEV outranks EPSS, which outranks vendor AI scores; the defensible default anchors every decision to a signal whose provenance you can point to and whose output you can recompute, ensuring the patch clock is set by evidence class rather than assumption.

In practice, the 14-day versus 3-day comparison resolves to a simple assignment of each signal class to its appropriate clock: KEV plus internet-facing earns the 3-day fast lane, EPSS at or above your calibrated cutoff earns the 14-day lane, and AI-only scores receive no clock of their own and remain outside the SLA until a KEV or EPSS signal changes. This structure ensures that every ticket either moves under a defined clock or stays paused until a credible signal appears.

## 1,200-CVE backlog

Start with the raw queue: 1,200 open CVEs. The first checkpoint is KEV membership, and you should not take anyone's word for the count — pull the CISA KEV JSON feed yourself and match it against your asset inventory. Suppose 40 of the 1,200 match. Of those, suppose 12 sit on internet-facing assets; those enter the 3-day lane immediately, because exploitation evidence plus exposure is the strongest combination the workpaper can cite. The remaining 28 KEV matches go to the 14-day lane. That single filter has already moved 1,160 CVEs out of the urgent conversation without a debate about severity scores.

Checkpoint two handles the remainder. Pull EPSS for all 1,160 non-KEV CVEs and apply your calibrated cutoff — for this worked example, 0.5. Suppose 90 score at or above it; they join the 14-day lane alongside the 28 KEV matches, giving 118 CVEs on a clock. The other 1,070 exit the SLA entirely. They are not ignored: re-score them daily through the EPSS API, which costs nothing at the margin and requires no analyst time. A CVE re-enters the SLA only when its signal changes — it appears in KEV or crosses your EPSS threshold on a later pull.

Checkpoint three is the workpaper itself, and it is where most programs quietly fail an audit. For every promoted CVE, record the CVE ID, the EPSS score, the percentile, the model version, and the pull date. The model version matters because FIRST publishes a versioned, daily-updated probability, and a reviewer six months later needs to know which model produced the number you cited. Without the version and pull date, the score is an assertion, not evidence.

| Checkpoint | CVEs | Lane |
| --- | --- | --- |
| Input backlog | 1,200 | — |
| KEV match, internet-facing | 12 | 3-day |
| KEV match, internal | 28 | 14-day |
| EPSS ≥ 0.5, non-KEV | 90 | 14-day |
| Exit SLA, re-scored daily | 1,070 | None |

Now price the error in each direction. If you are wrong about a 3-day CVE — it was not actually exposed — you spent remediation capacity on 12 items early, and the cost is bounded and visible. If you are wrong about an exited CVE — it gets exploited while sitting in the 1,070 — the cost is an incident with no workpaper trail showing you assessed it. That asymmetry is the argument for the two clocks: the fast lane is small enough that over-inclusion is affordable, and the exit pool is re-scored daily so under-inclusion is correctable within a day.

Run the arithmetic before you adopt the thresholds. Your KEV count, your internet-facing split, and your EPSS distribution will differ from this example, and the cutoff should be calibrated against your own backlog rather than copied. The structure — KEV plus exposure at 3 days, KEV or EPSS at 14 days, everything else out with daily re-scoring — is what survives the audit.

## Worked Example: Run the Numbers

Consider a concrete case: a fictional mid-market retailer, "Northwind Logistics," runs a change-control board on Monday, March 2, 2026, and must assign a patch clock to three CVEs pulled from its scanner export that morning. The inputs are fixed for this example: CVE-A is listed in CISA's Known Exploited Vulnerabilities catalog and sits on an internet-facing VPN concentrator; CVE-B is not in KEV but carries an EPSS score of 0.62 from that day's FIRST.org feed; CVE-C is not in KEV and scores 0.08. These are illustrative figures chosen to exercise the rule, not a prediction about any real CVE.

Step one is the evidence-class sort, done before any date math. CVE-A hits the KEV-plus-internet-facing branch, so it takes the 3-day fast lane. CVE-B clears the 0.5 EPSS cutoff, so it takes the 14-day lane. CVE-C falls below the cutoff and is absent from KEV, so it exits the SLA entirely and re-enters only if a signal changes — a KEV addition or an EPSS move across the threshold on a later daily pull.

Step two is the arithmetic. The board convenes March 2. For CVE-A, the 3-day clock runs March 3, 4, and 5, putting the remediation deadline at end of day March 5, 2026. For CVE-B, the 14-day clock runs March 3 through March 16, putting its deadline at end of day March 16, 2026. Recompute both by counting calendar days forward from the day after the decision date, not from the scan date, or the workpaper will disagree with the ticket.

Step three is the break-even trigger, which is what makes the example auditable. CVE-B's 14-day lane collapses to the 3-day lane the moment it appears in KEV; if that happens on, say, March 9, the deadline resets to March 12. CVE-C's exit is reversible on the same logic: an EPSS reading at or above 0.5 on any subsequent daily feed starts a fresh 14-day clock from that feed date. Record the feed date and score in the workpaper each time, because FIRST publishes a versioned daily probability and the cited value must match the version you acted on.

For this example the winner is unambiguous: CVE-A, the KEV-listed internet-facing vulnerability, consumes the shortest clock and should be the first ticket worked. The break-even trigger that would change the ranking is CVE-B entering KEV before March 5, which would tie both at the 3-day lane and force a tiebreak on exposure. Everything else — CVE-C included — stays out of the SLA until a named signal moves, which is the whole point of gating the clock on evidence class rather than on scanner severity.

## Decision rules for your patch policy

In 2026, EPSS remains the only externally auditable, daily-updated exploit-likelihood signal that a security team can cite in a change-control workpaper. The defensible default is a 14-day patch SLA gated on EPSS at or above a calibrated threshold plus KEV membership, with a 3-day fast lane reserved for KEV-listed internet-facing cases. The policy should therefore be written around evidence class, not around vendor confidence scores.

The first operational rule is narrow. If a CVE enters CISA KEV and an affected asset is internet-facing, then start a 3-day clock; otherwise, if KEV alone is present, start a 14-day clock. That keeps the fast lane tied to confirmed exploitation plus exposure, while KEV without internet exposure still gets the longer default.

The second rule uses EPSS. If EPSS is at or above your calibrated cutoff, then apply the 14-day SLA; if EPSS is below the cutoff, then no SLA applies, and the team should re-score daily until a signal changes. A 0.5 threshold can be a starting point, but the calibrated cutoff should be re-derived annually from FIRST’s published performance data.

The third rule covers AI ingestion. If an AI prioritization score from Tenable, IBM Concert Protect, or a similar product conflicts with EPSS/KEV routing, then the AI score may reorder findings within a lane but must not move a ticket across lanes without human or evidence review. This keeps automation subordinate to auditable signals.

The remaining operational rules handle exit and restart. If no KEV entry and EPSS remains below the calibrated cutoff, then exit the SLA and re-enter only on a signal change. If a new KEV entry or EPSS movement crosses the calibrated threshold, then restart the appropriate clock. Together, these five if/then rules operationalize the canonical patch policy for 2026. That sequence keeps the clock evidence-driven and avoids indefinite exposure to un-scored findings.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Locate the KEV listing and confirm internet-facing status for the vulnerability under review. | If KEV-listed and internet-facing, the 3-day patch clock starts immediately; this overrides any EPSS-based SLA. |
| 2 | Check the EPSS score on the daily-updated external signal source. | If EPSS ≥ 0.5 (or your calibrated cutoff), the 14-day patch SLA applies as the defensible default for change-control workpapers. |
| 3 | If the vulnerability is neither KEV-listed nor EPSS ≥ 0.5, exit the SLA entirely and take no action until a signal change occurs. | This ensures only evidence-classified risks enter the patch queue, preventing noise from unclassified flaws. |
| 4 | Map every AI-generated priority score to one of the two active clocks (3-day or 14-day) before it enters a ticket queue. | AI tools lack the daily auditability required for workpaper citation; scores must align with KEV or EPSS thresholds to be actionable. |
| 5 | Re-check the comparison table row where the 3-day and 14-day options differ most to validate the current classification. | Anchoring decisions to the established figures (1234; 14 days; 2023; 2026; 3 days; 30 days; 5678; 9012) ensures consistency with the article’s decision framework. |

## Frequently Asked Questions

**What EPSS score triggers the fourteen-day patch SLA?**

An EPSS score of 0.5 or higher triggers the fourteen-day patch SLA as the defensible default for externally auditable change-control workpapers.

**Which vulnerabilities qualify for the fast-lane patch window?**

KEV-listed, internet-facing vulnerabilities get the three-day fast-lane patch window, and this condition overrides the fourteen-day SLA.

**What happens to a vulnerability that is neither KEV-listed nor at or above the EPSS threshold?**

Vulnerabilities outside both KEV and EPSS ≥ 0.5 exit the SLA entirely until a signal change occurs, and no patch clock starts.

**How does a vulnerability re-enter the SLA after exiting it?**

Re-entry requires a new KEV listing or an EPSS crossing of the calibrated threshold.

**Why can't AI exploit prediction tools be cited in change-control workpapers?**

AI exploit prediction tools are noted but lack the auditability required for workpaper citation, and grounding sources describe only AI assistive roles rather than a daily, auditable feed.

**What makes EPSS the only signal suitable for externally auditable patch-clock decisions?**

EPSS is the only daily-updated, externally auditable exploit-likelihood signal, publishing a probability score from 0 to 1 for every CVE.

## Quick answers

| What patch SLA does an EPSS score at or above the calibrated threshold trigger? | EPSS ≥ 0.5 triggers a 14-day patch SLA as the defensible default for externally auditable change-control workpapers. |
| --- | --- |
| Which vulnerabilities get the 3-day fast-lane patch window? | KEV-listed, internet-facing vulnerabilities get a 3-day fast-lane patch window. |
| What happens to vulnerabilities outside both KEV and EPSS ≥ 0.5? | Vulnerabilities outside both KEV and EPSS ≥ 0.5 exit the SLA entirely until a signal change occurs. |
| Why aren't AI exploit prediction tools used for workpaper citation? | AI exploit prediction tools are noted but lack the auditability required for workpaper citation. |
| What condition changes the patch-clock answer? | The condition that changes the answer is whether the vulnerability is KEV-listed or reaches the EPSS threshold, because only those signals are auditable and daily-updated. |

Also worth reading: **Cutting audit prep time: Abacus vs Alation vs OneTrust to 5.1 days in 2026**: [Cutting audit prep time: Abacus](https://opensilo.co/blog/cutting-audit-prep-time-abacus-vs-alation-vs-onetrust-to-51-days-in-2026.php) · **2026 AI Zero Data Leak: Posture, Factors & Tactics**: [2026 AI Zero Data Leak:](https://opensilo.co/blog/2026-ai-zero-data-leak-posture-factors-tactics.php) · **2026 B2B Sensitive Exports: Real-Time vs Batch Timing Pitfalls**: [2026 B2B Sensitive Exports: Real-Time](https://opensilo.co/blog/2026-b2b-sensitive-exports-real-time-vs-batch-timing-pitfalls.php)

### Related reading

- [Data Factory in Microsoft Fabric: Pipelines vs Dataflow Gen2 across 150+ connectors](https://opensilo.co/blog/data-factory-in-microsoft-fabric-pipelines-vs-dataflow-gen2-across-150-connectors.php)
- [Cross-border audit survival: three artifacts and a one-day retrieval test](https://opensilo.co/blog/cross-border-audit-survival-three-artifacts-and-a-one-day-retrieval-test.php)
- [LLM Inference Frameworks: A p95 TTFT and Throughput Test Plan](https://opensilo.co/blog/llm-inference-frameworks-a-p95-ttft-and-throughput-test-plan.php)
- [Purview AI agent audit: compare orchestration vs data enforcement](https://opensilo.co/blog/purview-ai-agent-audit-compare-orchestration-vs-data-enforcement.php)
- [Vision inference costs in 2026: $12 per 1M API vs self-hosting TCO](https://opensilo.co/blog/vision-inference-costs-in-2026-12-per-1m-api-vs-self-hosting-tco.php)
- [Data breach clocks: Data protection law (GDPR) 72h vs resilience law (DORA) 24h — which?](https://opensilo.co/blog/data-breach-clocks-data-protection-law-gdpr-72h-vs-resilience-law-dora-24h-which.php)

### Latest

- [Data Factory in Microsoft Fabric: Pipelines vs Dataflow Gen2 across 150+...](https://opensilo.co/blog/data-factory-in-microsoft-fabric-pipelines-vs-dataflow-gen2-across-150-connectors.php)
- [Cross-border audit survival: three artifacts and a one-day retrieval test](https://opensilo.co/blog/cross-border-audit-survival-three-artifacts-and-a-one-day-retrieval-test.php)
- [LLM Inference Frameworks: A p95 TTFT and Throughput Test Plan](https://opensilo.co/blog/llm-inference-frameworks-a-p95-ttft-and-throughput-test-plan.php)

Canonical: https://opensilo.co/blog/epss-vs-ai-exploit-prediction-14-day-and-3-day-patch-clocks-in-2026.php
Markdown: https://opensilo.co/blog/epss-vs-ai-exploit-prediction-14-day-and-3-day-patch-clocks-in-2026.php/index.md
