The Complete Guide to Zero-Day Vulnerabilities
What a zero-day actually is, how the word is misused, how these flaws are found and sold, how disclosure works, and what defenders can do before a patch exists.

In December 2021 a flaw in Apache Log4j 2, a logging library almost nobody outside software engineering had heard of, turned into one of the most consequential vulnerabilities of the decade. Tracked as CVE-2021-44228 and scored 10.0 on the CVSS version 3.1 base scale by the National Vulnerability Database, it let an attacker who could get a string into a log message run code on the server. The library was embedded in so many products that a great many organizations could not answer the only question that mattered: do we use it?
Log4Shell was a zero-day for a short window. Then it stopped being one, and most of the damage happened afterward, to organizations that knew about it and had not finished patching. That pattern is the single most useful thing to understand about zero-days, and it is the opposite of how the term is usually discussed.
This guide explains what a zero-day actually is, how the word gets stretched beyond usefulness, how these flaws are found and traded, how disclosure is coordinated, and what a defender can realistically do during the window when no fix exists.
What zero-day actually means
The term comes from counting days since a patch became available. Zero days of patch availability means no fix exists. That is the whole idea.
Three related things get the same label, and separating them prevents most confusion.
- A zero-day vulnerability is the flaw itself, unknown to the vendor or known but unfixed, with no update available.
- A zero-day exploit is working code that takes advantage of that flaw. Plenty of vulnerabilities never get a reliable exploit, because the gap between "this looks wrong" and "this executes code every time" is substantial engineering work.
- A zero-day attack is the use of that exploit against real systems before a fix exists.
The clock stops when the vendor ships a fix. From that moment the flaw is a known vulnerability, often called an n-day, and the defensive question changes completely. For a zero-day you cannot patch, so you reduce exposure and watch for exploitation. For an n-day you patch, and anything else is a stopgap.
How the term gets misused
Four misuses are common enough to be worth naming.
Calling any unpatched system a zero-day. A server running software three years behind on updates is not exposed to a zero-day. It is exposed to dozens of known vulnerabilities with patches sitting in a repository. Describing it as a zero-day problem makes it sound like bad luck rather than unfinished work.
Calling a newly published CVE a zero-day. Most CVE records are published alongside or after a fix. That is coordinated disclosure working as intended, not a zero-day.
Treating a proof-of-concept as exploitation. Research code demonstrating a flaw is not evidence that anyone has attacked anyone. CISA is explicit on this point in its criteria for the Known Exploited Vulnerabilities catalog: scanning, security research of an exploit, and the public availability of a proof-of-concept do not themselves constitute active exploitation.
Marketing use. "Stops zero-day threats" appears in product material for tools that work by matching known patterns. Nothing can reliably detect an unknown exploit by its signature, because the signature does not exist yet. What good tools do is notice what happens after the exploit succeeds.
The vocabulary that is actually useful
If "zero-day" is doing too much work, these terms do the work precisely.
| Term | What it tells you |
|---|---|
| CVE ID | A unique identifier for one publicly known vulnerability, assigned by a CVE Numbering Authority. It is a name, not a judgment. A record can be Reserved, Published or Rejected. |
| CVSS base score | How severe the flaw is in the abstract, from the properties of the vulnerability itself. The specification is clear that the base score reflects severity, not the risk to your particular environment. |
| CISA KEV listing | That there is reliable evidence of exploitation in the wild, that a CVE ID exists, and that a clear remediation action is available. The strongest widely available signal that something is being used against real systems. |
| EPSS score | A daily probability between 0 and 1 that a published CVE will be exploited in the wild within the next 30 days, produced from observed activity rather than expert judgment. |
| Exploited in the wild | An attacker has used it against systems without permission. This includes attempted exploitation against systems that turned out not to be vulnerable. |
| Proof of concept | Code demonstrating the flaw exists. A step toward weaponization, not evidence of attack. |
| n-day | A known vulnerability with a patch available. Where most real-world compromise comes from. |
A sentence built from these is far more informative than one built from "zero-day". "CVE-2023-44487 was listed in KEV and the CVE record states it was exploited in the wild from August through October 2023" tells a defender what to do. "There is a new zero-day in HTTP/2" does not.
How zero-days are discovered
Zero-days are not found by accident as often as the word suggests. Most come from one of five routes.
Deliberate research. Someone reads code, reverse-engineers a binary, or runs automated fuzzing that feeds malformed input to software until it crashes, then investigates whether the crash is controllable. This is how most memory-corruption bugs in widely used software are found, and it is slow, specialized work.
Bug bounty programs. Vendors pay researchers for reports. This channel surfaces a large volume of findings and keeps many of them out of other hands, though the payouts for the highest-impact flaws rarely match what other buyers offer.
Variant hunting after a disclosure. When one bug in a component is published, researchers immediately look for the same pattern nearby. A disclosed vulnerability frequently produces a cluster of related ones over the following weeks, which is why a single advisory is often followed by several more in the same product.
Incident response. Some zero-days are discovered because somebody was already being attacked with them. An investigator finds an intrusion with no explained entry point, works backward, and identifies a previously unknown flaw. Vulnerabilities found this way are exploited in the wild by definition, which is why they tend to produce emergency patches.
Supply chain discovery. Occasionally the flaw is not a mistake. CVE-2024-3094 was malicious code discovered in the upstream source archives of the xz compression utility, present in versions 5.6.0 and 5.6.1, inserted through an obfuscated build-time mechanism and carrying a CVSS version 3.1 base score of 10.0. It was found because a developer investigated unusual performance behavior. That category is covered in our guide to software supply chain security.
The exploit market
A working exploit for widely deployed software has buyers, and understanding who they are explains a lot about how long flaws stay secret.
Vendors, through bug bounty programs, pay for reports so they can fix them. This is the only channel where the outcome is a patch.
Defensive vendors and researchers use findings to build detection, publish research and demonstrate capability. Disclosure usually follows.
Brokers buy exploits and resell them, typically to government customers, under terms that keep the flaw unpatched because a patched flaw is worthless to the buyer.
Government buyers acquire capability for intelligence and law enforcement purposes. Several jurisdictions run formal processes to decide whether to retain a vulnerability or report it to the vendor, and the fact that such processes exist tells you retention is a real outcome.
Criminal buyers operate the lower end for commodity flaws and the higher end for anything that enables ransomware deployment at scale.
Prices are driven by how widely deployed the target is, how much user interaction the exploit needs, how reliably it works, and how long it is likely to survive before being patched. A chain that compromises a fully updated phone with no user interaction is at the top because all four factors are favorable. Published price lists exist but are advertising, and most actual transactions are never confirmed, so treat any specific figure with caution.
The structural problem is that the buyer who wants the flaw fixed is usually not the highest bidder. That asymmetry is why coordinated disclosure depends partly on researcher ethics and partly on vendors making the reporting path easy and the treatment of reporters decent.
Disclosure models
Once someone knows about a flaw, they decide what to do with the knowledge. There are four broad answers, and the arguments between them have been running for decades.
Coordinated vulnerability disclosure is the mainstream model. The finder reports privately, the vendor acknowledges and develops a fix, both agree on a publication date, and details go public when the patch is available. A deadline is usually set so that an unresponsive vendor cannot stall indefinitely. NIST's SP 800-216 sets out recommendations for federal disclosure guidelines along these lines, the CERT Coordination Center publishes a detailed practitioner guide to the process, and CISA operates its own coordination program for flaws reported to it.
Full disclosure publishes everything immediately, on the argument that defenders deserve the same information attackers may already have and that public pressure is what gets some vendors to act. It is now uncommon as a first choice, and more often a response to a vendor refusing to engage.
Non-disclosure keeps the flaw private indefinitely, which is what the exploit market pays for.
Multi-party coordination is what happens when a flaw affects a component used by many vendors. FIRST publishes guidelines for exactly this situation, and it is genuinely hard: every additional party increases the chance of a leak, while leaving a party out means their customers are unprotected on publication day. Open-source components used across hundreds of downstream products are the worst case, because there is no central list of who ships them.
What happens on publication day
The release of a patch starts a race, and defenders usually lose the first lap.
- Patch and advisory published. The advisory describes the flaw well enough for administrators to assess it, which also tells researchers where to look.
- Reverse engineering of the patch. Comparing the fixed version to the previous one frequently reveals the flaw. For straightforward bugs this takes hours.
- Proof-of-concept publication. Working code appears publicly, often within days for internet-facing software.
- Mass scanning. Whole-internet scans for the vulnerable version begin almost immediately once an exploit is available.
- Mass exploitation. Anyone still unpatched is compromised indiscriminately rather than targeted.
This is why CISA's advisories on routinely exploited vulnerabilities keep reporting the same finding: attackers get most of their value from vulnerabilities that have had fixes available for some time. The window between patch and exploitation has compressed to days in many cases, and organizations whose patch cycle is measured in months are not in that race at all. Our guide to vulnerability management as a process covers how to close that gap.
What defenders can do before a patch exists
This is the part that gets skipped, usually because the honest answer is "less than you would like, but not nothing". The techniques below work because they do not depend on knowing what the exploit looks like.
Reduce what is reachable
Almost every mass-exploited zero-day of recent years has hit something that was reachable from the internet and did not need to be: a management interface, a file transfer appliance, a remote access gateway, a device administration portal. Every service you remove from the internet is a class of future zero-day that cannot touch you directly.
The practical version of this is an inventory of internet-facing services, reviewed on a schedule, with a named owner and a justification for each one. Anything without a justification gets restricted to a VPN, an allowlist of source addresses, or removed. This is unglamorous and it is the highest-value zero-day control available, because it works against vulnerabilities that have not been discovered yet.
Limit what exploitation gets the attacker
An exploited vulnerability on an isolated system with no stored credentials and no route to anything else is an incident. The same exploit on a system with domain administrator credentials cached and flat network access to everything is a catastrophe. The difference is architecture, decided long before the vulnerability existed. Network segmentation and the access model described in our zero trust guide are what convert the first outcome into the second being impossible.
Detect the behavior, not the exploit
You cannot write a signature for an unknown exploit. You can detect what attackers do next, because post-exploitation behavior is remarkably consistent across completely different initial access methods. A web server process starting a command shell, a sudden outbound connection from a machine that never initiates them, credential dumping from memory, new scheduled tasks, archive files appearing in temporary directories before a large upload. These are the detections that catch zero-days, and they are the same detections that catch everything else. Our threat intelligence guide covers how to build a detection list around behaviors rather than indicators.
Apply mitigations honestly
Vendors often publish a workaround before a patch: disable a feature, change a configuration, block a path. These are worth applying, with two caveats. A workaround is a temporary measure and should be tracked as such, with a ticket to transition to the real patch. And workarounds derived from early exploit samples are often incomplete, because a filter written against one exploitation path is frequently bypassed by the next variant. Web application firewall rules in particular should be treated as reducing exploitation volume rather than closing the hole.
Be ready to patch fast
The ability to deploy an emergency patch to an internet-facing system within hours is a capability you either have or do not, and you cannot build it during an incident. It requires knowing what you run, having a change process with an emergency path, having somewhere to test, and having an owner who is reachable. Organizations that patched Log4Shell quickly were almost always the ones that already had a dependency inventory, not the ones who worked harder that week.
Assume a breach you did not detect
For a serious internet-facing zero-day with evidence of exploitation before the patch, patching is not sufficient. The patch closes the door; it does not evict anyone who already came through. Several large incidents have involved organizations patching promptly and then discovering months later that an attacker had established access during the pre-patch window. If a vendor advisory says exploitation occurred before disclosure, the right response includes looking for signs of compromise, not only applying the update.
How to react when a zero-day is in the news
Most zero-day news does not require you to do anything, and the ones that do require a decision quickly. A short process keeps both from going wrong.
- Do we run it? Check the specific affected versions and the affected configuration, not the product name alone. Many advisories apply only to specific builds or to a feature that is off by default. If the answer is no, record that you checked and stop.
- Is it reachable, and by whom? An internal-only system with a vulnerability requiring network access is a very different priority from the same software published on the internet.
- Is there evidence of exploitation? Check whether it appears in the KEV catalog and what the vendor advisory actually says, rather than what coverage of it says.
- What does exploitation require? Authentication, user interaction, a specific configuration or local access each change the picture substantially.
- What can we do today? Patch if a fix exists. If not, restrict access, apply the vendor workaround, and raise monitoring on the affected systems.
- What would exploitation look like in our logs? Decide this before you need it, and confirm you are collecting the logs that would show it.
Note where the official prioritization thinking has moved. CISA's binding directive on prioritizing security updates based on risk, which carries forward the KEV criteria established by its 2021 predecessor, sets federal remediation timelines from a combination of whether the flaw is in KEV, how automatable exploitation is, the technical impact, and whether the affected asset is publicly exposed. The most urgent category requires action within three days alongside a forensic triage of the asset to establish whether it has already been compromised. The specific deadlines apply to US federal agencies, but the model of combining exploitation status with exposure is a sound way for anyone to decide what to do first.
Who actually needs to worry about targeted zero-days
Being honest about this prevents wasted effort. An expensive zero-day gets spent on targets worth the cost of burning it, because using an exploit risks its discovery and the end of its usefulness. That realistically means government systems, critical infrastructure, large technology companies holding other organizations' data, journalists, dissidents and senior executives at significant commercial targets.
For most organizations the relevant exposure is not being targeted with a zero-day. It is being caught in the indiscriminate wave that follows public disclosure, where attackers scan broadly and compromise whoever has not patched. The defenses for that are patching speed, reduced internet exposure and the ability to notice intrusion, all of which are the same investments you would make anyway. The small-organization version of this program is covered in our small business cybersecurity guide.
Zero-day news often arrives as a phishing hook. "Critical security update, install immediately" messages spike after any major vulnerability story. Paste a suspicious message or link into CyberWatch AI Scan for a free check, and get patches only from the vendor's own site.
Mistakes worth avoiding
- Buying a product because it claims to stop zero-days. Ask what it detects after exploitation and how, then judge that claim instead.
- Treating a high CVSS score as an instruction. The base score describes the flaw, not your exposure to it. A 9.8 in software you do not run from the internet can rank below a 7.5 in your public login page.
- Patching and closing the ticket on a vulnerability that was exploited before the patch existed, without looking for evidence that someone got in first.
- Letting a workaround become the permanent answer. Track mitigations with an expiry, because mitigations get bypassed and forgotten.
- Ignoring the boring internet-facing appliance. File transfer products, remote access gateways and management portals have repeatedly been the entry point for mass exploitation, and they are often owned by nobody in particular.
The short version
A zero-day is a flaw with no available fix, and the label expires the moment the vendor ships one. The word is applied far more widely than that, usually to unpatched known vulnerabilities, which have a completely different answer.
True zero-days are found by deliberate research, bought by people who benefit from them staying unfixed, and disclosed through a coordination process that works better than its critics admit and worse than its defenders claim. During the window with no patch, what helps is having less exposed to the internet, less reachable from each compromised system, and better detection of what attackers do after they get in.
And when the patch arrives, the thing that decides your outcome is how quickly you can apply it. That is where the real risk has always been.
Frequently asked questions
What is a zero-day vulnerability?
A flaw in software or hardware that the vendor does not yet have a fix for at the time it is being used or disclosed. The name comes from counting the days a patch has been available: zero. Once a fix exists, the flaw is no longer a zero-day, even if most systems remain unpatched.
What is the difference between a zero-day vulnerability and a zero-day exploit?
The vulnerability is the flaw itself. The exploit is the working code that takes advantage of it. A vulnerability can exist for years with no exploit, and the existence of an exploit is what turns a theoretical problem into an operational one.
Is an unpatched vulnerability on my systems a zero-day?
No. If the vendor has released a fix and you have not applied it, that is an unpatched known vulnerability, sometimes called an n-day. The distinction matters because the answer to a known vulnerability is to patch it, and the answer to a true zero-day is to reduce exposure and watch for exploitation.
Do zero-days cause most breaches?
No. The large majority of intrusions use known vulnerabilities with available patches, stolen credentials, phishing or misconfiguration. Zero-days matter disproportionately in targeted attacks against well-defended organizations, which is a real but much smaller share of incidents.
How much is a zero-day worth?
Prices range from a few thousand for a low-impact bug through a vendor bug bounty, to very large sums for a reliable chain that compromises a fully updated phone without any user interaction. Published figures vary widely and should be treated as indicative, since the market is private and most transactions are never confirmed.
What is coordinated vulnerability disclosure?
The practice of reporting a flaw privately to the vendor, agreeing a timeline for a fix, and publishing details once a patch is available or the timeline expires. It balances giving the vendor time to fix against the public's need to know, and it is the model most national cybersecurity agencies and researchers follow.
What can I do about a vulnerability with no patch?
Reduce exposure by removing the affected service from the internet or restricting who can reach it, apply any vendor workaround, make sure you can detect the behavior that follows exploitation rather than the exploit itself, and prepare to patch quickly when a fix lands.
What does exploited in the wild mean?
That there is reliable evidence of an attacker using the flaw against real systems without permission. It is a much stronger statement than a proof-of-concept existing or a vulnerability being scanned for, and it is the threshold CISA uses for its Known Exploited Vulnerabilities catalog.
Should small organizations worry about zero-days?
Only indirectly. Small organizations are rarely the target of an expensive zero-day, but they are frequently caught in the mass exploitation wave that follows a disclosure, when attackers scan the internet for anyone who has not patched yet. Speed of patching matters far more than zero-day defenses.
Does my antivirus or EDR stop zero-days?
Not by signature, because there is nothing to match yet. Modern endpoint tools can still catch the activity that follows a successful exploit, such as a web server spawning a command shell or credentials being dumped from memory. That behavioral detection is the realistic defense, and it is partial.
Sources
- Glossary, CVE Program
- Process, CVE Program
- Reducing the Significant Risk of Known Exploited Vulnerabilities, US Cybersecurity and Infrastructure Security Agency
- BOD 26-04: Prioritizing Security Updates Based on Risk, US Cybersecurity and Infrastructure Security Agency
- CVSS v4.0 Specification Document, FIRST
- Exploit Prediction Scoring System (EPSS), FIRST
- Guidelines and Practices for Multi-Party Vulnerability Coordination and Disclosure, FIRST
- The CERT Guide to Coordinated Vulnerability Disclosure, Carnegie Mellon University CERT Coordination Center
- SP 800-216, Recommendations for Federal Vulnerability Disclosure Guidelines, US National Institute of Standards and Technology
- 2023 Top Routinely Exploited Vulnerabilities, US Cybersecurity and Infrastructure Security Agency


