Blog

EDR and XDR: what each one sees, what it misses, how to choose

How EPP, EDR and XDR work, what telemetry each collects, where they tend to fail and what to ask before choosing one for your company.

A closed laptop next to a rack server

Antivirus, EDR and XDR are sold as rungs on a ladder, and the buying conversation tends to stop at the product name. What matters to your company is something else: what data the tool collects, which machines it is actually installed on, what it does on its own when it sees something suspicious, and who looks at the alert when the automatic response is not enough. This article explains how each layer works and where each one tends to fail.

EPP, EDR and XDR: three different questions

An EPP (Endpoint Protection Platform) answers the question “should this file or process run?”. It is what antivirus grew into: signatures, file reputation, machine learning models that classify binaries before they run, and behavior rules that block known patterns, such as an Office document launching powershell.exe. The EPP acts before and during execution, and decides on its own.

An EDR (Endpoint Detection and Response) answers a different question: “what happened on this machine, in what order, and how do I stop it?”. An agent collects continuous telemetry from the operating system: process creation with the full command line, parent and child processes, network connections, file writes, registry changes, module loads, access to other processes’ memory. These events go to a console where rules and models look for suspicious sequences, and where an analyst can rebuild the story of an attack. The response side includes isolating the machine from the network, killing processes, quarantining files and, in some products, rolling back changes made by ransomware.

An XDR (Extended Detection and Response) widens the question to “what happened in the environment?”. It joins endpoint telemetry with other sources, such as identity (Active Directory, Microsoft Entra ID), cloud workloads, email and network, and correlates events that look harmless on their own. An unusual sign-in to Entra ID, followed by an odd rundll32 run on a server and a bulk LDAP query, can become a single incident, if the sources are integrated, instead of three loose alerts in different consoles.

What the attacker does and what the tool must see

MITRE ATT&CK is the most useful way to think about coverage, because it describes real techniques observed in intrusions. A few that a well-configured EDR should see:

  • Execution through interpreters (T1059): PowerShell with a Base64-encoded command, or wscript launched by an attachment. mshta has its own technique (T1218.005).
  • Credential theft from memory (T1003.001): reading the memory of lsass.exe, whether with known tools or with Windows’ own comsvcs.dll writing a dump.
  • Lateral movement (T1021): admin shares or RDP with stolen credentials, and remote execution through WMI (T1047).
  • Impairing defenses (T1685): trying to stop the agent itself, turn off event logging (T1685.001) or add exclusions.
  • Legitimate tools: attackers prefer binaries already on the system (certutil, bitsadmin, commercial remote access tools), because that use is harder to tell apart from normal administration.

That last point is why an EPP alone is not enough. When an attacker uses only what is already installed, there is no malicious file to block. What remains is behavior, and behavior can only be analyzed with telemetry and context.

Where EDR and XDR tend to fail

Often the failure is not in the detection engine but in deployment and tuning:

  • Incomplete coverage. Legacy servers, supplier machines, VMs created by scripts, and Linux or macOS endpoints without an agent. Those are exactly the machines an attacker looks for.
  • Policies left in detect-only mode. The product was installed to watch and never moved to blocking, for fear of stopping operations.
  • Broad exclusions. Whole folders, or backup and old antivirus processes, excluded from analysis, giving the attacker a safe place to work.
  • Tamper protection turned off. A local administrator can uninstall the agent.
  • Noise. Too many alerts about legitimate internal tools teach the team to ignore the console, and the real alert gets lost.
  • Nobody responds. The tool detects, but there is no defined process for who acts, how fast, and with what authority to isolate a production machine.

XDR adds one more risk: integrations that exist in the contract but were never connected. An XDR that only receives endpoint data is an EDR under another name.

An illustrative example

This scenario is a composite of findings common in penetration tests, not a real customer’s case.

A retailer has EDR on every workstation and on most servers. An employee receives an email with an .iso file; inside it, a shortcut runs a script that downloads a loader. The EDR detects the execution and raises a medium-severity alert, but the policy is in detect-only mode for the finance team’s workstations, because an old reconciliation system kept triggering false positives. The alert sits in the queue.

With access to the workstation, the attacker queries Active Directory and finds a service account with an old password and admin rights on several servers. The attacker cracks the password offline (Kerberoasting) and uses the account to reach a file server that has no agent, because it was migrated from another data center and dropped out of the inventory. From there on, what happens on that server is no longer recorded by any agent.

What would have changed the outcome: active blocking for the finance group, with a narrow exclusion for the reconciliation binary; a cross-check between the directory and the console to find machines without an agent; and an identity layer able to flag the unusual Kerberos ticket requests. The first two need no product change; the identity layer may need an integration or a license. All of them need configuration, and someone who knows the attacker’s path.

How to choose between EDR and XDR

EDR is the mandatory starting point: without endpoint telemetry, an investigation loses almost everything that happened on the machines. XDR makes sense once your team runs the EDR consistently and the bottleneck has become joining up what happens in identity, cloud and email. If the team cannot yet handle the alerts it receives, more data sources will not help.

Before deciding, measure three things in your environment: how many assets have a working agent compared with the real inventory, how many policies are in blocking mode, and how long a critical alert takes to be seen by a person. Those three answers say more about your protection than the product’s acronym.

What to ask a supplier

  1. What telemetry does the agent collect on Windows, Linux and macOS, and what differs between them?
  2. How does the product handle techniques that use only native system tools, such as an LSASS dump with comsvcs.dll?
  3. Which sources does the XDR actually integrate in my environment today, and which need an extra license or connector?
  4. How does tamper protection work, and can a local administrator remove the agent?
  5. Who sets and reviews the exclusions, and how are they documented?
  6. After deployment, who tunes the policies, and who responds to a critical alert outside business hours?

How Leonnes works with EDR and XDR

Leonnes does not make EDR or XDR. We work with SentinelOne for endpoints, cloud and identities: we assess what your company already has, test the product in your environment before you decide, deploy it, tune policies and alerts with an attacker’s eye, and give your team technical support in day-to-day operation. Learn more about Leonnes’ Blue Team.

See what an attacker would see first

We reply within one business day.