The Patchless Window

The days between a Public Code Fix and the CVE that instructs defenders to patch.

Open source security fixes land as public commits before the CVE is published. AI agents can now read those commits at scale and identify new vulnerabilities in order to create working exploits. Until the CVE is published, defenders have no signal to patch while AI-accelerated attackers work undisturbed.

Days measured from the Public Code Fix to CVE publication.

The Patchless Window: from the Public Code Fix to CVE publication only attackers can act; after publication the familiar race begins. EMERGING · KNOWN · THE PATCHLESS WINDOW · From Public Code Fix · THE FAMILIAR RACE · From CVE publication · Attackers work undisturbed · An exploit can be built from the Public Code Fix in hours. · Defenders have no signal to patch · No CVE means no scanner finding, no ticket and no patch cycle. · Their edge fades · as patches roll out · Slow start · then they catch up as they patch · ATTACKERS · DEFENDERS · Public Code Fix · CVE publication · Median 39 days in August 2026 · Days to months after publication

Attackers’ advantage, stronger where solid Defenders’ protection, stronger where solid Nothing to act on
Figure 1. The Patchless Window vs The Familiar Race.
39 days
The Patchless Window Median, August 2026
18,594
Open source CVEs measured January to August 2026
1.8 days
Fastest decile of public exploits After the Public Code Fix

The race everyone knows starts at CVE publication

For most of the industry, a patch lifecycle begins when its CVE is published. A vendor or project announces that version X is affected and version Y fixes it; the record appears in public CVE databases, and the race is on. Defenders start finding vulnerable systems and deploying patches. Meanwhile, attackers are busy weaponizing the flaw.

This race has been measured for years, and the attacks are coming faster than ever. The Zero Day Clock, a public dashboard of vulnerability exploitation, tracks the time from disclosure to exploitation in the wild. Its data shows a significant decrease in that time, from about a year in 2021 to just over a day in 2026.

In the first days after publication, attackers hold the advantage; as patches roll out over the following weeks, it shifts to defenders. But the race no longer waits for publication: Mandiant’s M-Trends 2026, published in March, puts mean time to exploit at an estimated minus seven days, meaning exploitation now routinely begins before a patch is even released. Since March, the picture has gotten worse. This study adds an earlier point to the familiar race, the moment the fix itself becomes public, and measures how far ahead of publication it now comes: from about a week in March to more than a month by August.

But the clock starts at the Public Code Fix

In open source, that fix is public: a commit anyone can read, often days or weeks, and in some cases months, before publication. We refer to it as the Public Code Fix. The commit message frequently names the bug class (e.g., XSS, buffer overflow) and sometimes the impact as well (e.g., remote code execution, privilege escalation). Sometimes it goes further, with hints about how the flaw is reached and exploited. Some even reference a GitHub Security Advisory (GHSA) ID, which GitHub assigns as soon as a maintainer drafts an advisory. That draft can already hold a registered, not-yet-public CVE ID, so the commit marks a security fix while the CVE is still considered private. Some Public Code Fixes ship with tests that trigger the bug, to validate the fix. Running them against the unpatched version is essentially the exploit itself. The difference between vulnerable and fixed code is, in effect, a specification for the exploit.

Defenders cannot follow every commit in every product or dependency. They wait for the CVE, as do the scanners and vulnerability management platforms that tell them what to patch. Chainguard brought the point home when it disclosed 14 Java vulnerabilities that had been fixed upstream with no identifier: “A fix with no identifier is a fix only the attacker who read the commit knows about.”

The time between the Public Code Fix and CVE publication is what we call The Patchless Window. Within that interval, AI-accelerated attackers have sufficient information for exploitation, while defenders usually are not aware there is a vulnerability.

We measure the window from the Public Code Fix to CVE publication. A silent patch, a fixed version already released by the open source project, may ship earlier, but almost nobody knows to install it while there is still no CVE. CVE publication is what defenders and vulnerability management tools act on.

A related term already exists. An n-day vulnerability is one whose fix is already public, with n counting the days since that fix became available. It stays exploitable wherever the fix hasn't been deployed. A vulnerability inside The Patchless Window is, in effect, a CVE-less n-day. What sets it apart is that defenders have not been told: the fix is public, but no CVE exists. That is why this study counts from the date of the Public Code Fix, not from the date of the CVE.

Eight moments in a vulnerability’s life

  1. 1
    Vulnerable code introduced
    The flaw enters the codebase.
  2. 2
    CVE registration
    A CVE ID is reserved, not yet public.
  3. 3
    Public Code Fix STRONG SIGNAL FOR AI
    The fix is public and points at the flaw. The Patchless Window opens.
  4. 4
    Silent patch
    A fixed version ships, but no CVE calls defenders to action.
  5. 5
    CVE publication
    The flaw and the official patch are announced. The window closes.
  6. 6
    Patch deployment
    Defenders find and update vulnerable systems.
POSSIBLE MOMENTS AFTER STEP 1
  • Public exploit
    Working exploit code is published.
  • KEV listing
    Exploitation in the wild is confirmed.

These can land anywhere after step 1 or not at all.

Figure 2. The vulnerability lifecycle.

If you can read the source, AI can write the exploit

This is the thesis behind the study, stated plainly: in 2026, access to source code is close to access to an exploit. Open source publishes the vulnerable code, the fix and the exact changes between them, under version control and timestamped to the minute. A capable model needs nothing more to start.

Attackers have always read patches. What changed is who can do it, and at what scale. Reading the code changes, understanding the bug and writing an exploit used to take a skilled researcher days to months of focused work; models now do much of it, across thousands of repositories at once. DARPA’s AI Cyber Challenge showed where this was heading back in 2025, when autonomous systems analyzed 54 million lines of open source code and found 18 real vulnerabilities.

Open source hands over the answer key, and the new models turn reading it into a routine task, at a scale no human team could match. Closed-source software brings a second challenge: attackers have to reverse-engineer and compare compiled binaries, which is still not as easy as reading a commit, though the latest models have made major progress there as well.

Public source
+
The code changes
+
Capable AI Agent
=
A working exploit
Figure 3. The premise behind this study: once an open source fix is public, building a working exploit becomes far easier.

How we measured The Patchless Window

To measure the window, we worked backward from disclosure. Taking the open source CVEs we monitored (published from January to August 2026), we went to the public code repositories. For each CVE, we found the repository that holds the affected code, and inside that repository pinpointed the first commit that fixed the flaw: its Public Code Fix.

  1. 01
    Gather the CVEs
    We gathered all open source CVE IDs published from January to August 2026.
  2. 02
    Find the repository
    For each one, we identified the public code repository that holds the affected code.
  3. 03
    Pinpoint the Public Code Fix
    We went back commit by commit and analyzed each one as a possible lead to the vulnerability, assessing its vulnerable code flows and exploitability. We then verified each lead against our proprietary vulnerability analysis built for that CVE. We identified the first commit that fixed the vulnerability and traced its first public appearance: the first moment anyone outside the project could have seen it.
  4. 04
    Measure The Patchless Window
    We measured the exact time between that Public Code Fix and the CVE’s publication. A silent patch may come earlier, but the window ends at publication, when the fix becomes something to act on.
Figure 4. The Patchless Window measurement process.

This resulted in two timestamps per CVE: when the fix became public, and when the CVE did. Working at scale, we identified the Public Code Fix for 18,594 CVEs. For exploit timing, we matched public exploits from the VulnCheck Exploit Database (XDB) and other public exploit repositories to the same Public Code Fixes and measured from the commit to the exploit’s first public appearance.

In 2026, The Patchless Window grew from 5 days to 39

We matched 18,594 CVEs published from January to August 2026 to their Public Code Fix and measured the gap. In January, the median CVE was published 5.1 days after its Public Code Fix. By August it was 39.2 days, across 3,786 CVEs published that month.

Median days from Public Code Fix to CVE publication by month, January to August 2026: 5.1, 7.9, 6.4, 12.3, 20.0, 28.7, 36.8, 39.2. 0 · 10 · 20 · 30 · 40 days · 5.1 · January · 7.9 · February · 6.4 · March · 12.3 · April · 20.0 · May · 28.7 · June · 36.8 · July · 39.2 · August · 2026

Figure 5. Median days from Public Code Fix to CVE publication, by month, January to August 2026.

The growth is measured. Its cause is not, so what follows are hypotheses, not findings. The first is volume: in 2026, AI began finding vulnerabilities at industrial scale. Anthropic alone, using Claude Mythos Preview, has disclosed 6,157 vulnerabilities across 591 open source projects since November 2025.

Others are scaling the same way. OpenAI’s Daybreak program has reported findings of its own, including more than 400 privilege escalation flaws in a single operating system kernel.

The growth is sharpest in open source. Black Duck’s 2026 OSSRA report found that the average number of open source vulnerabilities per codebase more than doubled in a year. Many of those findings become reports for CVE Numbering Authorities to process, and that load may be part of why the delay between the Public Code Fix and publication is growing.

We are not forecasting that the window will continue to widen. We are reporting that in 2026 it widened, and that every extra day is a day attackers can work while defenders are largely unaware there is anything to act on, and their tools, which key on CVEs, can do little without one. Shortening it, which can be done in more than one way, is a shared job for maintainers, CNAs, security vendors and security teams.

Public exploits arrived within hours of the fix

If the Public Code Fix is the specification, how quickly does someone build the exploit? Unit 42 already describes threat actors “automating the ‘monitor → diff → test → weaponize’ loop.” Back in 2022, Unit 42 reported that “attackers typically start scanning for vulnerabilities within 15 minutes of a CVE being announced.” To see how fast that loop runs in public, we measured the time from the commit to the exploit’s first public appearance. These are only the exploits their authors chose to publish. With today’s tools, we believe many more stay private, and the published ones were likely ready before they were uploaded.

The five examples below are not the five fastest overall. They were selected from the fastest decile, among widely known products. In each case, exploit code went public within a day of the Public Code Fix, sometimes just a few hours later. Not all of it came from attackers: in some cases, the researchers who reported the flaw, or the maintainers who fixed it, published a proof of concept soon after the fix. Whoever wrote it, the exploit was public. The CVE, and with it the defenders’ starting signal, came days or weeks later.

#CVEProductPublic Code Fix → exploitPatchless Window
1CVE-2026-54088filebrowser4h 2m22 days
2CVE-2026-75604Next.js4h 31m7 days
3CVE-2026-46300Linux kernel5h 12m10 days
4CVE-2026-34200nhost8h 24m5 days
5CVE-2026-34227Sliver16h 40m5 days
Table 1. Public Code Fix to first public exploit, whoever published it, and each CVE’s Patchless Window. Five widely known products from the fastest decile. Exploit data: VulnCheck Exploit Database (XDB) and other public exploit repositories.

The same pattern shows up in widely exploited high-severity vulnerabilities. Cisco Talos put it plainly: “Proof-of-concept code for exploiting new vulnerabilities used to take attackers months to create. Now they take hours.”

Taken together: the fastest exploits went public within hours of the fix, while The Patchless Window’s median was 39 days in August 2026. That gives attackers a decisive advantage. An attacker who lacks the skills to write an exploit sometimes does not need to: they can download one when available, and still have plenty of time before defenders are told to patch.

The fastest tenth of exploits arrived in 1.8 days

To look past a handful of examples, we ranked every CVE in our exploit data from the shortest time between Public Code Fix and public exploit to the longest and split the list into ten equal groups, or deciles. The first decile is the fastest tenth of all public exploits. The chart shows the five fastest deciles (half of the data set). The exploit timings cover only CVEs with a public exploit: 884 of them, about 88 in each decile, a much smaller set than the 18,594.

Days from Public Code Fix to public exploit for the five fastest deciles: 1.8, 6.9, 14.1, 22.8 and 34.4 days. 0 · 10 · 20 · 30 · 40 days · The Patchless Window, August 2026 median: 39 days · 1st decile · (fastest) · 1.8 days · Attackers’ head start · 2nd decile · 6.9 days · 3rd decile · 14.1 days · 4th decile · 22.8 days · 5th decile · 34.4 days

Figure 6. CVEs from 2025 and 2026 with a public exploit, ranked by time from Public Code Fix to exploit and split into ten equal groups; the five fastest are shown. Exploit data: VulnCheck Exploit Database (XDB) and other public exploit repositories.

The first decile sits at 1.8 days and the fifth at 34.4 days. All five are shorter than The Patchless Window’s August 2026 median of 39 days. In other words, for approximately half of the CVEs with a public exploit, the exploit went public in less time than the August 2026 median Patchless Window. And that counts only exploits someone chose to publish.

A note on the comparison. The 39-day median is the window for CVEs published in August 2026. The exploit timings come from a broader set of CVEs from 2025 and 2026 that have a public exploit. The two come from different groups of CVEs, so the median is a reference point.

We expect the deciles to compress. As models improve and access to them widens, the slower deciles should move toward the first. Work on the defending side is growing too. Anthropic’s Cyber Mission, announced in October 2026, includes OSS Scanner, an opt-in service that scans enrolled open source projects and reports findings with suggested fixes. Initiatives like it may help maintainers find and fix flaws sooner, which could slow the trend.

If the window stays measured in weeks while exploitation settles at hours or days, attackers enjoy a head start of weeks on every vulnerability fixed in public.

Three fixes, three windows

Case 1 · CVE-2025-59057 · React Router

Fixed in September. Announced to defenders in January.

A fixed version shipped two days later, with no sign it closed a vulnerability. For four months, until the CVE, defenders had no reason to install it.

Timeline for CVE-2025-59057, drawn to scale. THE PATCHLESS WINDOW · 121 DAYS · Public Code Fix · 0 · Silent patch · +42h 23m · CVE publication · +121d · Public exploit · +124d

Case 2 · CVE-2026-48778 · Notepad++

A public exploit, 29 days before the CVE.

The fix and a silent patch shipped two minutes apart, and the commit already named the CVE ID. A public exploit followed two days later.

Timeline for CVE-2026-48778, drawn to scale. THE PATCHLESS WINDOW · 31 DAYS · Public Code Fix · 0 · Silent patch · +2m · Public exploit · +2d · CVE publication · +31d

Case 3 · CVE-2026-75604 · Next.js

An exploit in under five hours. The CVE a week later.

A silent patch followed the fix within two hours, and a public exploit within five. For seven days, exploit code was public while scanners had no CVE to flag.

Timeline for CVE-2026-75604, drawn to scale. THE PATCHLESS WINDOW · 7 DAYS · Public Code Fix · 0 · Silent patch · +1h 38m · Public exploit · +4h 31m · CVE publication · +7d

Figure 7. Three Patchless Windows, from Public Code Fix to CVE publication, drawn to scale.

Where the window really opens and closes

We drew The Patchless Window from the Public Code Fix to CVE publication because those are the two points in time we can measure across thousands of CVEs. Both ends deserve a second look.

The flaw exists from the moment vulnerable code is merged, so exposure starts before the Public Code Fix.

At the other end, exposure does not stop when the CVE is published. In most organizations, the CVE record triggers the scanner finding, the ticket, and the patch cycle, and systems are fixed days or weeks later. Every day of that sequence comes on top of a window that has already been open for weeks. A defender’s exposure ends when they are protected (a deployed patch or compensating control), which can come long after publication.

The real gap runs from vulnerable code to protection, and it should shrink from both ends. Finding flaws closer to the moment they were written pulls its start toward the Public Code Fix. Protection that arrives closer to the Public Code Fix pulls its end the other way. As AI shortens the path from a public fix to a working exploit, the time from the Public Code Fix to protection is likely to become a defining measure of a security program, one teams should track as closely as time to patch. The more both ends close in, the better protected the ecosystem is.

Top: the Patchless Window as measured, from the Public Code Fix to CVE publication. Below it: the real gap, from vulnerable code introduced to protection, and the desired gap, its two ends closing in from both sides. THE PATCHLESS WINDOW · Public Code Fix · CVE publication · Median 39 days, August 2026 · THE REAL GAP · Vulnerable code introduced · Protected · THE DESIRED GAP · Goal: shrink the time from vulnerability signals to protection

Figure 8. Top: the window we measured. Bottom: the real gap, from vulnerable code to protection, and the smaller gap to aim for.

The balance is clear. A CVE publication in many cases helps attackers more than defenders. For an attacker, it is a pointer: go hunt for a public exploit, or generate one. For defenders, it is the start of a race to protect, against a clock already running before they received any signal.

We will return to both ends of the window in future research. Until then, every second counts.

Stay in the know

Get our latest research, news and updates.

Privacy Policy

Last modified October 11, 2026

Introduction

This Privacy Policy outlines Second Security Inc.'s ("we", "our" or "the Company") practices with respect to information collected from users who access our website (the "Site") or communicate with us in any other way (collectively: "Users").

We encourage Users to carefully read the Privacy Policy and use it to make informed decisions.

In this Privacy Policy you can read about:

  • What type of information we collect
  • How we use and with whom we share the information
  • Our Legal basis for processing your information
  • User rights
  • For how long we retain the information
  • How we safeguard the information
  • How to contact us

What Information we collect?

We collect two types of data and information from Users who access our Site.

The first type of information is un-identified and non-identifiable information pertaining to a User(s), which may be made available or gathered via your access to the Site (“Non-personal Information”). We are not aware of the identity of a User from which the Non-personal Information was collected. Non-personal Information which is being collected may include your aggregated usage information and technical information transmitted by your device, including certain software and hardware information (e.g. browser type, language preference, access time, etc.) in order to enhance the functionality of our Site. We may also collect information on your activity on the Site (e.g. clicks, actions, etc.).

The second type of information is individually identifiable information, namely information that identifies an individual or may with reasonable effort identify an individual (“Personal Information”). Such information includes:

  • Contact information: the Site may provide you with several options to fill in contact information. If you choose to contact us through our Site or by any other means you will be asked to provide us with your name, email address and, in some cases, your phone number.
  • Contact information: When you visit or log in to our website, cookies and similar technologies may be used to associate these activities with other personal information they or others have about you, including by association with your email.
  • Device Information: We may also collect Personal Information from your device. Such information includes geolocation data, IP address, unique identifiers (e.g. MAC address and UUID) and other information which relates to your activity through the Site.

If we associate Non-personal Information with Personal Information, we will treat such information as Personal Information.

How is the information used? With whom do we share the information?

We do not rent, sell, or share Users’ information with third parties except as described in this Privacy Policy.

We may use the information for the following:

  • Communicating with you – If you have contacted us, for any reason, we will use your information in order to further communicate with you.
  • Conducting statistical and analytical purposes, intended to improve the Site.

In addition to the different uses listed above, we may transfer or disclose Personal Information to our subsidiaries, affiliated companies and subcontractors.

In addition to the purposes listed in this Privacy Policy, we may share Personal Information with our trusted third-party providers, who may be located in different jurisdictions across the world, for any of the following purposes:

  • Providing you with our Service, including providing a personalized display of our Site;
  • Storing and processing such information on our behalf;
  • Performing research, technical diagnostics or analytics;

We may also disclose information if we have good faith to believe that disclosure of such information is helpful or reasonably necessary to: (i) comply with any applicable law, regulation, legal process or governmental request; (ii) enforce our policies, including investigations of potential violations thereof; (iii) investigate, detect, prevent, or take action regarding illegal activities or other wrongdoing, suspected fraud or security issues; (iv) to establish or exercise our rights to defend against legal claims; (v) prevent harm to the rights, property or safety of us, our users, yourself or any third party; or (vi) for the purpose of collaborating with law enforcement agencies and/or in case we find it necessary in order to enforce intellectual property or other legal rights.

In any case we will conclude that disclosure of Personal Information is necessary, we will only disclose that portion of the Personal Information that is required in order to satisfy the basis for disclosure. We will share the Personal Information under appropriate safeguards.

Updates from Second Security

If you sign up on the Site to receive updates, we collect the email address you provide. We use it only to send you research, news and updates related to Second Security, and we do not sell it or share it for any other purpose.

Sign-ups are received and stored on our behalf by Basin (usebasin.com), the form service the Site uses for this purpose. Basin processes your email address only to deliver it to us.

You can stop receiving updates at any time by writing to us at privacy@second.security, and we will remove your email address from our list. We keep your email address for as long as you remain subscribed, and delete it once you unsubscribe, unless we are required by law to retain it.

User Rights

Depending on the jurisdiction in which you reside, you may have certain rights under relevant applicable laws regarding the collection and processing of your Personal information. To the extent these rights apply and concern you, you can contact us via the contact details available below and ask us to exercise the following rights:

  • Receive confirmation as to whether or not personal information concerning you is being processed, and access your stored personal information, together with supplementary information.
  • Receive a copy of personal information you directly volunteer to us in a structured, commonly used and machine-readable format.
  • Request rectification of your personal information that is in our control.
  • Request erasure of your personal information.
  • Object to the processing of personal information by us.
  • Request to restrict processing of your personal information by us.
  • Lodge a complaint with a supervisory authority.

However, please note that these rights are not absolute, and may be subject to our own legitimate interests and regulatory requirements.

If you wish to exercise any of the aforementioned rights, or receive more information, please contact us at privacy@second.security.

Retention

We will retain your personal information for as long as necessary to provide our services, and as necessary to comply with our legal obligations, resolve disputes, and enforce our policies. Retention periods will be determined taking into account the type of information that is collected and the purpose for which it is collected, bearing in mind the requirements applicable to the situation and the need to destroy outdated, unused information at the earliest reasonable time. Under applicable regulations, we will keep records containing client personal data, account opening documents, communications and anything else as required by applicable laws and regulations.

We may rectify, replenish or remove incomplete or inaccurate information, at any time and at our own discretion.

Third party collection of information

Our policy only addresses the use and disclosure of information we collect from you. To the extent you disclose your information to other parties or sites throughout the internet, different rules may apply to their use or disclosure of the information you disclose to them. Accordingly, we encourage you to read the terms and conditions and privacy policy of each third party that you choose to disclose information to.

This Privacy Policy does not apply to the practices of companies that we do not own or control, or to individuals whom we do not employ or manage, including any of the third parties which we may disclose information as set forth in this Privacy Policy.

How do we safeguard your information?

We take great care in implementing and maintaining the security of the Site and your information. We employ industry standard procedures and policies to ensure the safety of the information we collect and retain, and prevent unauthorized use of any such information, and we require any third party to comply with similar security requirements, in accordance with this Privacy Policy. Although we take reasonable steps to safeguard information, we cannot be responsible for the acts of those who gain unauthorized access or abuse our Site, and we make no warranty, express, implied or otherwise, that we will prevent such access.

Transfer of data outside the EEA

Please note that some data recipients may be located outside the EEA. In such cases we will transfer your data only to such countries as approved by the European Commission as providing adequate level of data protection, or enter into legal agreements ensuring an adequate level of data protection.

Corporate transaction

We may share information in the event of a corporate transaction (e.g. sale of a substantial part of our business, merger, consolidation or asset sale). In the event of the above, the transferee or acquiring company will assume the rights and obligations as described in this Privacy Policy.

Minors

The Site is intended for the use of individuals over the age of 18. In the event that we become aware that a user under the age of 18 has shared any information, we will discard such information. If have any reason to believe that a child under 18 has shared any information with us, please contact us at privacy@second.security.

Updates or amendments to this Privacy Policy

We reserve the right to periodically amend or revise the Privacy Policy; material changes will be effective immediately upon the display of the revised Privacy policy. The last revision will be reflected in the "Last modified" section. Your continued use of the Platform, following the notification of such amendments on our website, constitutes your acknowledgment and consent of such amendments to the Privacy Policy and your agreement to be bound by the terms of such amendments.

How to contact us

If you have any general questions about the Site or the information we collect about you and how we use it, you can contact us at privacy@second.security.

Zvi Strahilevitz St 4, Tel Aviv-Jaffa