Wazuh SIEM implementation

Wazuh SIEM implementation

... typical errors & alarm levels!

A mid-sized network with Windows and Linux servers, a firewall, and Microsoft 365 generates thousands of log entries every day. No one reads them manually.

A Security Information and Event Management (SIEM) system centrally collects this data, normalizes it, and evaluates it based on rules so that the few events indicating a potential attack stand out from the mass.

Wazuh is an open-source platform that you can self-host without licensing costs. However, a Wazuh instance left unchanged after installation primarily generates noise.

Most deployments fail due to a lack of tuning and undefined processes that no one has established beforehand. Both can be planned before the rollout.

How Wazuh is structured

Wazuh consists of four components. The agent runs on Windows, Linux, and macOS endpoints, where it reads logs, inventory data, and file changes, then sends them to the Wazuh server. The active connection uses port 1514/TCP, while new agent registration uses port 1515/TCP.

The server decodes the events and checks them against the rule set. Matches are stored by the Wazuh indexer, a search and analytics cluster, and the Wazuh dashboard makes them evaluable in the browser. In version 4.x, Filebeat transports the alerts from the server to the indexer, by default over port 9200/TCP with TLS.

For devices without an agent, such as firewalls, switches, or routers, the server receives syslog. The listener on port 514 is disabled by default; you enable it in the server’s ossec.conf. The agent reads Windows event logs directly via the Event Channel, monitoring the System, Application, and Security channels by default. Additional channels like Sysmon or PowerShell are added as a separate <localfile> block. Winlogbeat is not required for this.

Microsoft 365 integrates with Wazuh via two modules. The office365 module reads activities from the Microsoft 365 tenant, while the ms-graph module retrieves security alerts and incidents from Microsoft Defender as well as Intune audit events via the Graph API. If you also want to fetch Entra sign-in logs via the Graph API, Microsoft requires an Entra ID P1 or P2 license in the tenant.

Wazuh 5.0 is available as a Release Candidate and does not yet have a GA date. The new major version overhauls the architecture: Filebeat is removed, the server writes directly to the indexer, and initial tests from the community show that custom rules from 4.x need to be adjusted.

For an introduction, this means: Start with 4.14 and plan the migration to 5.0 as a separate project once the version is generally available.

Related to this: Entra ID Admin Center: Monitoring and Integrity shows you which sign-in and audit logs Entra ID generates so you can decide which ones should be integrated into the SIEM before connecting, and Why P1 or P2 for Security explains what additional features the Premium license includes, which is required for the Graph API retrieval of sign-in logs.

additional Wazuh features

Agents are configured centrally via agent groups. A new log source or additional monitoring path is thus deployed to all hosts in the group without having to touch each server individually. Three modules provide information that goes beyond classic log management.

Vulnerability detection

The agent inventories the operating system and installed packages via the Syscollector module. The server compares this list with the Wazuh Cyber Threat Intelligence (CTI). The CTI aggregates data from vendors such as Microsoft, Red Hat, Ubuntu, and Debian, as well as from NVD, OSV, and the CISA catalog.

If an installed version falls within the affected range of a CVE, Wazuh generates a finding. On Windows, Wazuh additionally checks whether an installed hotfix already closes the vulnerability. A patched system thus no longer appears as vulnerable after patch day.

Bewertung der Sicherheitskonfiguration

The SCA module checks endpoint configurations against stored policies, which are primarily based on CIS Benchmarks. Issues like an enabled SMBv1 or disabled audit logging are flagged before an attacker can exploit them. Since misconfigurations can be exploited without any code vulnerabilities, it’s worth regularly addressing SCA findings rather than just monitoring the percentage score in the dashboard.

Dateiintegritätsüberwachung

FIM monitors specified files and directories and reports any changes. After a breach, an attacker typically modifies binary files, autostart entries, or configurations. Therefore, a FIM alert on a server where no one is currently working is an early warning sign. With Whodata enabled, the alert additionally includes the user and the process responsible for the change.

Sizing and Scaling

Overview by agent count (90-day retention)

AgentsDeploymentServer Nodes (vCPU, RAM)Indexer Nodes (vCPU, RAM)Total NVMe Storage
1 to 25All-in-One, 1 Host4 vCPU, 16 GiB (combined)on the server host100 GB
26 to 50All-in-One, 1 Host8 vCPU, 16 GiB (combined)on the server host150 to 200 GB
51 to 100All-in-One, 1 Host8 vCPU, 24 to 32 GiB (combined)on the server host250 to 300 GB
101 to 250small cluster2 × (8 vCPU, 8 GiB)3 × (8 vCPU, 16 GiB)600 GB to 1 TB
251 to 500medium cluster2 × (8 vCPU, 16 GiB)3 × (8 vCPU, 32 GiB)1.2 to 2 TB
501 to 1,000large cluster3 × (8 to 16 vCPU, 16 GiB)3 to 5 × (16 vCPU, 32 to 64 GiB)2.5 to 5 TB
1,001 to 2,500Enterprise cluster4 × (16 vCPUs, 16 GiB)5 to 7 × (16 vCPU, 64 GiB)6 to 12 TB

Storage requirements depend on event rate and retention period. If you need to keep alarms for 180 days instead of 90, roughly expect double the space. These figures are guidelines for the initial setup. After a few weeks of operation, you’ll see the actual event rate per second in the dashboard and can adjust accordingly.

Roles in a distributed cluster

This breakdown shows how a cluster for 100 to 500 agents is composed by machine roles:

RoleNodevCPU per nodeRAM per nodeDisk per node
Wazuh server (Master)188 to 16 GiB50 to 100 GB SSD
Wazuh server (Worker)1 to 388 GiB50 to 100 GB SSD
Wazuh Indexer (Data and Cluster Manager)minimum 38 to 1616 to 32 GiB300 to 800 GB NVMe
Wazuh Dashboard1, or on the Master44 to 8 GiB30 GB SSD
Load Balancer (NGINX or HAProxy)1 to 222 to 4 GiB20 GB SSD

Why the indexer needs at least three nodes

The Wazuh Indexer is based on OpenSearch. Every decision in the cluster, such as selecting the cluster manager, requires a majority of the voting-eligible nodes. With two nodes, that means both. If one fails, there is no longer a majority, and the remaining node will not accept write operations.

While a split-brain scenario prevents OpenSearch via this very majority rule, high availability is still lost with just two nodes. Only with three or more nodes does a majority of two remain after a failure. At the same time, the cluster can then secure every primary shard with a replica on a different node.

The 31 GiB limit for JVM heap

You configure the indexer heap in /etc/wazuh-indexer/jvm.options. Wazuh recommends allocating half of the available RAM, so for 32 GiB RAM:

-Xms16g
-Xmx16g

Never set the heap above 31 GiB, even if the host has 64 GiB of RAM. Beyond approximately 32 GiB, Java disables Compressed Ordinary Object Pointers (Compressed OOPs). Each object pointer then occupies 8 bytes instead of 4, the heap holds less data at the same size, and garbage collection becomes more expensive.

The second half of the RAM is intentionally left free: The Linux kernel uses it as a filesystem cache for the Lucene segments, which directly benefits searches in the dashboard. To prevent the operating system from swapping out the heap, you also set bootstrap.memory_lock: true in the opensearch.yml.

Load balancer in front of the server nodes

With two or more Wazuh server nodes, agents should be evenly distributed and automatically fail over to another node if one goes down. To achieve this, you place a network load balancer such as NGINX or HAProxy in front of the servers. It distributes event connections on port 1514/TCP across the worker nodes. Agent registration on port 1515/TCP is forwarded to the master node, or agents can connect directly to the master for this purpose.

The server nodes coordinate with each other via port 1516/TCP: The master distributes agent keys, agent groups, its own rules and decoders, SCA policies, and CDB lists to the workers over this port. If a worker fails, its agents are rerouted via the load balancer to the remaining nodes, and analysis continues.

Reading alarm levels correctly

wazuh alarm levels

Wazuh assigns each rule a level from 0 to 16.

This level indicates how security-relevant a single event is and thus determines storage and notification.

The
official classification is as follows:

LevelMeaning according to WazuhExample
ignoredRule suppresses known false alarms
1 to 2low relevance, system messagesStatus messages without security relevance
3successful or authorized eventssuccessful login, allowed firewall connection
4Low-priority system errorMisconfiguration, unused device
5User-induced errorWrong password, denied action
6Low attack relevanceVirus with no effect, frequent IDS events
7Keyword matchMessage contains 'bad' or 'error'
8first observedfirst user login, first IDS alert
9error from invalid sourcelogin attempt by unknown user, error on admin or root account
10multiple user errorsrepeated incorrect passwords
11integrity check warningmodified binary files, rootkit detected
12High-priority eventSystem or kernel errors and warnings
13Unusual high-priority errorMatches a common attack pattern
14High-priority security eventUsually triggered by correlation
15Severe attackFalse alarm virtually impossible
16maximum severitynot used by any default rule, reserved for custom rules

Two thresholds in the ossec.conf determine what happens with these levels. log_alert_level defaults to 3, anything below that does not generate an alert. The email threshold email_alert_level is set to 12 in the included manager configuration, whereas the reference documentation lists 7 as the default. Check the value in your installation before relying on email notifications.

The severity level only provides the basis for escalation—you define the thresholds yourself. A good starting point is this approach: Monitor alarms up to level 7 in the dashboard for trend analysis. Review levels 8 to 10 daily, as they often indicate new logins and brute-force patterns. Address level 11 the same day, since modified binaries may signal a successful compromise.

Starting at level 12, a notification is sent to the on-call team, and levels 14 to 16 trigger your incident response workflow. Adjust the thresholds once you see, after a few weeks of operation, how many alerts are assigned to each level.

Related: Admin rights only when truly needed explains how to reduce permanent admin rights using PIM, making alerts for privileged accounts less frequent and each one more meaningful.

Common mistakes during implementation

No target and no use cases

"We need a SIEM because compliance requires it" is not a project goal. Before installation, define which attacks you want to detect: brute force attacks against the VPN, logins from countries where no one in your company operates, changes to the Domain Admins group. These use cases will determine the necessary data sources and rules.

Conversely, those who first collect all logs and then think about what to do with them will fill up storage and generate alerts that no one can classify.

Determine in the same step whether cloud services, Active Directory, or Linux servers pose your greatest risk. This will decide which systems you connect first and how much hardware you’ll need in the initial months.

Not enough time for operations

A SIEM is not a product you set up once and forget. Someone needs to read, classify, and escalate alerts when necessary—and this exact effort is often missing from project plans. After deployment, schedule dedicated time for rule tuning, data source maintenance, and alert analysis. Without it, the rule set becomes outdated, and the team gets used to dismissing alerts with a click.

Gaps in data sources

A SIEM can only detect what it sees. If a critical system’s log is missing, that’s exactly where a blind spot remains. Ideally, you integrate these sources:

  • Network devices such as firewalls, switches, and routers
  • Servers and clients running Windows, Linux, and macOS
  • Cloud services including Microsoft 365, Azure, AWS, and Google Cloud
  • Security components like IDS/IPS, antivirus, and Endpoint Detection and Response
  • Application logs from databases, web servers, and ERP systems

Endpoints send their data via the agent, network devices via Syslog to the server. If an application provides a log format that Wazuh doesn’t recognize, you write your own decoder. Without a decoder, the lines do end up in the system, but the rules can’t parse fields like user or source IP, so correlation fails.

Related: Defender Licenses: P1, P2, XDR & Business Premium explains which Defender plan actually generates EDR alerts and incidents that the ms-graph module can then forward to the SIEM.

Default rules without tuning

The default rule set of Wazuh is a starting point but doesn’t know your environment. Running it unchanged either generates a flood of false positives or misses events that are only critical in your network. The result is alert fatigue: the team ignores warnings because most of them are meaningless, and real attacks get lost in the noise.

Three levers help here. You define thresholds to specify how many failed logins within a given time window trigger an alert. Exceptions are set as child rules with level 0, for example for test systems or lab networks where high alert volumes are normal. Correlation rules group multiple similar events from different sources and increase their relevance. Custom rules are added in /var/ossec/etc/rules/local_rules.xml with IDs starting from 100000 to prevent them from being overwritten by updates to the standard rule set.

If your company works remotely a lot, VPN logins and cloud access deserve their own rules. In other environments, off-hours bookings are the critical pattern. In both cases, this separates false alarms from real threats.

Related to this: Conditional Access: Device Code Phishing Policy describes a specific attack pattern in Microsoft 365 for which you can write your own detection rule, and Enterprise Apps: Disable User Consent shows which consent events in the audit indicate OAuth phishing and therefore deserve an alert.

Lack of training

An alert is only useful if someone understands it. Those working with Wazuh need foundational knowledge in four areas: log analysis to identify malicious patterns, alert correlation to see connections between events, forensics to investigate systems under suspicion and secure evidence, and attack techniques to classify new patterns.

The MITRE ATT&CK framework helps with the last point, as Wazuh already maps many rules to ATT&CK techniques. Since attack methods are constantly evolving, continuous training should be a fixed part of the annual plan.

Outdated versions and rules

Wazuh releases patch versions in the 4.14 series approximately monthly in 2026: 4.14.6 on July 1, 4.14.7 on July 29, 4.14.8 on September 23, 4.14.9 on October 14, and 4.14.10 on November 5. Version 4.14.8, among other things, hardened path validation and login limits for the server API. Anyone who leaves such updates unattended for months is running a security tool with known vulnerabilities.

In addition to the software, you maintain your own indicators of compromise (IoCs). Wazuh ingests them as CDB lists, such as those containing known malicious IP addresses or file hashes, and matches incoming events against them.

No incident response workflow

A SIEM provides alerts, but no response. To turn a critical alert into a countermeasure, you first need answers to three questions: Who analyzes the alert, and who approves the countermeasures? What steps follow, such as a forensic image or immediate system isolation? Which channel do you use to communicate if your own network is compromised?

Wazuh can execute simple responses on its own. For example, Active Response blocks an attacking IP address on the affected host as soon as a specific rule triggers. For more complex workflows, you integrate a SOAR platform (Security Orchestration, Automation and Response). Shuffle is already included as an integration.

No documentation

Over time, every Wazuh installation accumulates its own rules, decoders, exceptions, and new data sources. Without proper documentation, troubleshooting takes longer, and knowledge becomes siloed with one person. That’s why you should store rules, decoders, and configurations in a Git repository. This way, every change gets a timestamp and a reason, and new team members can clearly see why a rule is structured the way it is.

further points ...

Compliance only claimed

Wazuh provides mappings of its rules to PCI DSS, GDPR, HIPAA, NIST 800-53, and the Trust Services Criteria (TSC), including corresponding dashboards. The documentation clarifies that these mappings are a guideline and do not replace formal certification. There are no pre-defined mappings for ISO 27001 or NIS2. However, you can add your own compliance identifiers to your rules and filter based on them.

This has been relevant for companies in Germany since December 6, 2025: On that day, the NIS2 Implementation Act came into force. Section 30 (2) of the BSIG requires affected entities, among other things, to implement measures for handling security incidents and assessing the effectiveness of their risk management. A SIEM provides evidence for both, but does not replace an information security management system.

Remember data protection. Logs contain usernames and IP addresses, which are personal data. Therefore, define retention periods and implement them in the indexer via policies for automatic deletion, rather than storing alert data indefinitely.

Integration with third-party systems

The Wazuh integrator service provides ready-made connections to Slack, PagerDuty, VirusTotal, Shuffle, and Maltiverse. To connect additional destinations—such as a ticketing system or a Teams channel—you use a custom script, with its name starting with custom-. This way, a critical alert automatically creates a ticket, and follow-up no longer depends on someone actively monitoring the dashboard.

The VirusTotal integration checks the hashes of files reported as new or modified by FIM against the VirusTotal database. Together with the CDB lists and Maltiverse, it supplements your own rules with external threat intelligence.

Community & Support

For questions about Wazuh, the community provides answers in the Slack workspace, the Google Groups mailing list, on Reddit, Discord, and in GitHub Discussions. There you’ll also find developer responses regarding release plans, such as the status of Wazuh 5.0. If you need guaranteed response times, you can purchase professional support, consulting, or training from Wazuh. Additionally, Wazuh Cloud offers a hosted solution where you don’t have to manage servers and indexers yourself.

What makes economic sense

Wazuh doesn’t require a license, but it does demand hardware and man-hours. The payoff comes when an attack is detected early enough to contain it before it cripples production or administration. Add to that the audit evidence you’d otherwise painstakingly gather from individual systems, and the relief for your team when automatic correlation replaces manual log searches.

Conclusion

Wazuh is a serious option for companies without their own Security Operations Center, provided someone is allocated dedicated time for tuning and analysis. Start with a few clearly defined use cases, classify alert levels according to the official taxonomy, write your own rules with exceptions and thresholds, and define the escalation path before the first Level-12 alarm hits.

Deploy version 4.xx today and only plan the switch to 5.0 once it’s generally available and you’ve tested your own rules against it.

Read more: Entra ID Admin Center: Conditional Access shows how you can block many of the sign-in attacks that Wazuh later reports as alerts in advance using conditional access policies.

Share:
No comments yet

Be the first to start the discussion with a helpful comment.

Leave a comment

Comments are reviewed briefly before they are published — factual, respectful and to the point is just right here.

Your email address will not be published.