Blog/Vulnerabilities

Antivirus, EDR and XDR: What Endpoint Protection Really Does

What antivirus, EDR and XDR each actually detect, the gaps every endpoint tool has, and the device hardening that matters more than the product you choose.

CyberWatch AI7 November 2026 · 7 min read
An open laptop and a smartphone side by side on a wooden desk

An analyst reviewing an alert sees a chain of events on a finance laptop: a spreadsheet application started a scripting host, the script reached out to a hosting service, and something was written to a folder that spreadsheets never write to. No file on that machine was flagged as malicious. Every program involved was legitimate and signed. Antivirus had nothing to say, because nothing bad was present as a file.

That gap between "is this file malicious" and "is this sequence of actions normal" is the whole reason endpoint security has three acronyms instead of one. Understanding what each actually does, and does not do, is more useful than any product comparison.

The three generations, side by side

Antivirus / EPPEDRXDR
Core questionIs this file or process malicious?What has been happening on this device?What has been happening across our estate?
Main methodSignatures, heuristics, reputation, machine learning on file featuresContinuous recording of process, file, registry and network activity, with behavioral detectionThe same, correlated with identity, email, cloud and network telemetry
OutputBlocked or allowedAlerts with a reconstructable timelineAlerts joining events across sources
ResponseQuarantine the fileIsolate the device, kill processes, collect artifacts, run remote commandsThe same, plus actions in other systems such as disabling an account
Needs a person?Largely noYes, or a managed serviceYes, or a managed service
Catches the example above?UnlikelyOftenOften, with more context

Two points about this table matter more than the columns.

First, these are not three tiers where each replaces the last. EDR products include the preventive engine; they add recording and investigation on top. You are not choosing between blocking and detecting.

Second, EDR and XDR have a prerequisite that antivirus does not: somebody has to look at the output. An EDR deployment with nobody reviewing detections is an expensive antivirus with a detailed log of an intrusion nobody noticed. This is the most common way endpoint spending gets wasted, and it is why managed detection services exist.

What XDR means in practice

The idea behind XDR is sound and worth stating plainly: individually weak signals become strong when joined. A sign-in from an unusual location is weak on its own. So is a process reading browser credential stores. Together, on the same account within ten minutes, they are an incident.

The difficulty is that "XDR" has no agreed definition. Some products mean an endpoint tool that also ingests email and identity logs from the same vendor. Some mean a platform aiming to replace a security information and event management system. Some mean a dashboard. When evaluating one, ignore the label and ask which data sources it correlates, whether it will take yours, and what detections exist that genuinely require more than one source.

The detection gaps every endpoint tool has

Vendors talk about coverage. Gaps are more informative, and these are structural rather than specific to any product.

Legitimate tools used maliciously. Attackers increasingly avoid custom malware in favor of what is already installed: scripting engines, administration utilities, remote support software, archive tools. The binary is signed, expected and often on an allowlist. Detection has to rest on unusual use rather than the program's identity, which produces both misses and false positives.

Credential theft and reuse. Once an attacker has valid credentials, much of what follows is indistinguishable from an administrator working. Logging into another machine with a stolen password is a normal event, and the abnormality lives in the pattern across many machines rather than on any one of them.

Devices with no agent. Network appliances, printers, building systems, cameras, industrial controllers and personal devices cannot run an endpoint agent. Several large intrusions have begun on exactly these, precisely because nothing watches them. Coverage is often quoted as a percentage of managed devices, which quietly excludes the whole category.

Deliberate interference. Disabling or modifying security tools is a recognized adversary technique, documented in the ATT&CK framework and observed constantly in real incidents. An attacker with administrative rights will try to stop the agent, exclude a folder from scanning, or uninstall the product. Tamper protection, central reporting of agent health and an alert when an agent stops checking in are the countermeasures, and the last one is the one most often missing.

Audit mode. Many deployments run with the strongest controls in detect-only mode after an early false positive broke something. Months later nobody remembers, and the product is reporting on attacks rather than stopping them. It is worth checking which of your protections are actually enforcing.

None of this is an argument against endpoint tooling. It is an argument for pairing it with controls that do not depend on correctly identifying malice, which is what hardening does.

Hardening: the part that works without detection

MITRE ATT&CK publishes mitigations alongside techniques, and reading them together is clarifying, because many techniques are addressed by configuration rather than by detection. The NCSC's device security guidance and the CIS Benchmarks give concrete settings per platform. The following are the changes that remove the most attacker options per unit of disruption.

Remove standing administrative rights from everyday accounts. A large share of what an attacker wants to do next requires local administrator privilege. Separation of user and administrator privilege appearing on the NSA and CISA list of most common misconfigurations tells you how often this is still unaddressed. Give people a second account or an elevation mechanism for the tasks that genuinely need it.

Control what can execute. Application control, where only approved software runs, is the strongest single endpoint control and the hardest to introduce. Start where the software set is stable and the stakes are high: servers, finance machines, administrative workstations. Even a partial deployment in enforcement mode beats a comprehensive one in audit mode.

Disable the macro and script paths you do not need. Macros in documents from the internet, and scripting hosts on machines that have no use for them, account for a persistent share of initial access. Our guide to macro security warnings covers what users see.

Patch the operating system and the browser automatically. On end-user devices, automatic updates with a rollback plan beat a controlled monthly cycle, because the exposure window is where the risk lives. Our vulnerability management guide covers how to prioritize everything else.

Encrypt the disk and require a screen lock. This addresses the most ordinary loss event there is, and it is covered for phones in phone screen locks.

Protect credentials on the device. Enable the platform features that isolate credential material in memory, discourage saving passwords in browsers in favor of a password manager, and remember that infostealer malware targets browser credential stores and session cookies specifically.

Make devices report their state centrally, so you can answer which machines are unpatched, unencrypted or missing an agent. Unmanaged devices with access to company data are the gap that makes every other control partial.

Phones are a different problem

Mobile operating systems isolate applications from each other, which removes most of what desktop endpoint products are built to detect. Scanning apps on a phone has limited value. What matters instead is enrollment in device management, keeping the operating system current, restricting installation to official stores, separating work data so it can be wiped without touching personal data, and a strong screen lock. CISA's mobile communications guidance covers the higher-risk end, including messaging choices for people likely to be targeted.

Choosing and deploying without wasting money

A short sequence of questions settles most of this.

  1. Who will read the alerts? If the answer is nobody, buy a managed detection and response service instead of a product, or stay with well-configured built-in protection until you have the capacity.
  2. What does it cover? Windows, macOS, Linux servers, cloud workloads and your actual mix of devices. Partial platform coverage is the most common unpleasant discovery after purchase.
  3. How long does it keep telemetry? Investigations routinely need weeks of history. Short retention turns a thorough investigation into guesswork.
  4. Can we get our own data out? You want the telemetry available to your own analysis, not only inside the vendor's console.
  5. What happens when the agent is tampered with? Confirm there is tamper protection and that a missing agent raises an alert.
  6. How will we test it? Running a set of known benign test behaviors mapped to ATT&CK techniques tells you far more than a vendor demonstration.

Then deploy in a specific order. Get the agent onto everything before tuning anything, because coverage gaps matter more than false positives. Turn protections to enforcement mode deliberately and record any exclusion you create with a reason and a review date. Set up the alert that fires when a device stops reporting. And rehearse isolating a machine once, so that nobody is reading documentation during an incident.

Think you already ran something you should not have? Paste the file name, download link or message into CyberWatch AI Scan for a free check, and disconnect the device from the network while you look.

What to remember

Antivirus asks whether a file is bad. EDR records what happened so a human can tell. XDR joins that recording to the rest of your environment. Each generation addressed a real gap in the one before, and none of them closed it completely, because attackers moved to living off legitimate tools and valid credentials where the question "is this malicious" has no clean answer.

The decisions that matter most are not which vendor. They are whether every device runs the agent, whether it is in enforcement mode, whether someone reads the alerts, whether users hold administrative rights they do not need, and whether you would notice an agent going silent. For how endpoint protection fits into the broader picture of malicious software, see our guide to malware and ransomware.

Frequently asked questions

What is the difference between antivirus and EDR?

Antivirus decides whether a file is malicious and blocks it, mostly before or as it runs. EDR records what processes actually do on the device, looks for suspicious sequences of behavior, and lets an analyst investigate and respond remotely. Antivirus answers whether something is bad; EDR answers what happened.

What does XDR add?

XDR correlates endpoint telemetry with other sources such as identity, email, cloud and network data, so a weak signal on a device can be joined to a weak signal elsewhere. The concept is sound and the term is used loosely, so what any given product means by it varies a great deal.

Is built-in antivirus good enough?

For an individual or a very small organization with no one to monitor alerts, a well-configured built-in product with its cloud protection features enabled is a reasonable position. Once you have systems worth investigating an intrusion on, the recording that EDR provides becomes the thing you are missing.

Does EDR stop ransomware?

It frequently detects the preparation stages, such as credential theft, backup deletion and mass file access, and can isolate the device automatically. It is not a guarantee, because attackers test against common products and because detection only helps if somebody acts on the alert.

Do Macs and phones need endpoint protection?

Macs benefit from the same detection and response capability as Windows machines and are increasingly targeted. Phones work differently: the operating system isolates apps, so the useful controls are device management, keeping the OS updated, restricting app installation sources and screen locks rather than a scanner.

Can attackers disable endpoint protection?

Yes, and it is a documented and common step once they have administrative rights. This is why tamper protection, an agent that reports its own health centrally, and alerting when an agent goes quiet all matter as much as the detection itself.

What is application control?

A policy that allows only approved software to run, instead of trying to identify bad software. It is one of the strongest endpoint controls available and one of the most disruptive to introduce, which is why it is usually applied to servers and high-risk roles first.

How many endpoint products should we run?

One detection product, configured properly and monitored. Running two often causes conflicts, performance problems and gaps where each assumes the other is handling something, and it rarely improves coverage.

What matters more, the tool or the configuration?

The configuration and whether anyone reads the alerts. A mid-market product in blocking mode with tamper protection on and a person reviewing detections daily outperforms a market-leading product left in audit mode with nobody watching.

Sources

  1. SP 800-83 Rev. 1, Guide to Malware Incident Prevention and Handling for Desktops and Laptops, US National Institute of Standards and Technology
  2. Techniques - Enterprise, MITRE ATT&CK
  3. Mitigations - Enterprise, MITRE ATT&CK
  4. Data Sources, MITRE ATT&CK
  5. Device security guidance, UK National Cyber Security Centre
  6. Device security guidance | Managing deployed devices, UK National Cyber Security Centre
  7. CIS Benchmarks, Center for Internet Security
  8. NSA and CISA Red and Blue Teams Share Top Ten Cybersecurity Misconfigurations, US Cybersecurity and Infrastructure Security Agency
  9. Mobile Communications Best Practice Guidance, US Cybersecurity and Infrastructure Security Agency
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.