kql-mde-xdr

name: kql-mde-xdr description: Write, optimize, and debug KQL (Kusto Query Language) queries for Microsoft Defender for Endpoint (MDE), Microsoft Sentinel, and Microsoft 365 Defender XDR. ALWAYS use this skill when the user mentions DeviceProcessEvents, DeviceFileEvents, DeviceNetworkEvents, DeviceLogonEvents, DeviceRegistryEvents, AlertInfo, or any MDE/Sentinel table names. Use for threat hunting queries, detection rules, incident investigation, IOC hunting, MITRE ATT&CK detections, query optimization, or converting SPL/other query languages to KQL. Trigger on phrases like "write a KQL query", "defender query", "sentinel query", "hunt for", "detection rule", "M365 defender", or any security analysis involving Microsoft security products. refs: - references/*.md

KQL and MDE XDR Expert

Core Capabilities

1. KQL Query Writing - Craft optimized queries for security analysis
2. Threat Hunting - Proactive hunting queries for adversary techniques
3. Detection Engineering - Create detection rules and analytics
4. Incident Investigation - Investigate alerts and incidents
5. Performance Optimization - Optimize slow or inefficient queries

MDE Tables Reference

CRITICAL: You MUST read the relevant reference files BEFORE writing any query. Do not rely on memory - always verify field names, ActionTypes, and table schemas from the reference documentation.

Query Writing Principles

Use Native Fields First

Before using string manipulation functions like split(), extract(), or parse(), check if the data you need already exists as a native field in the table schema. For example:

• Use FileName directly instead of extracting from FolderPath
• Use InitiatingProcessFileName instead of parsing command lines
• Check AdditionalFields for structured data before regex parsing

Focus on the Right Entity

When detecting suspicious activity, consider which entity matters most for detection:

• Persistence detection: Focus on InitiatingProcess* fields (what created/modified the persistence)
• Malware detection: Focus on the file/process itself (FileName, FolderPath)
• Lateral movement: Focus on source/remote fields

KQL Fundamentals

Query Structure

TableName
| where TimeGenerated > ago(24h)
| where <condition>
| project <columns>
| summarize <aggregation> by <grouping>
| order by <column> desc

Essential Operators

String Functions

// Case-insensitive contains
| where FileName contains "mimikatz"

// Case-sensitive contains
| where FileName contains_cs "Mimikatz"

// Starts/ends with
| where FileName startswith "cmd" or FileName endswith ".ps1"

// Regex matching
| where FileName matches regex @"(?i).*mimi.*"

// String extraction
| extend Domain = extract(@"https?://([^/]+)", 1, Url)

Time Functions

// Relative time
| where Timestamp > ago(7d)
| where Timestamp between (ago(48h) .. ago(24h))

// Time binning
| summarize count() by bin(Timestamp, 1h)

// Time formatting
| extend Hour = datetime_part("hour", Timestamp)

Query Optimization

Performance Best Practices

1. Filter early - Use where clauses first to reduce dataset size
2. Time bound - Always include time filters
3. Use has over contains - has is faster for word boundaries
4. Avoid * in project - Select only needed columns
5. Limit join scope - Filter tables before joining

Inefficient vs Optimized

// INEFFICIENT - late filtering, uses regex
DeviceProcessEvents
| project-away ReportId
| extend lower_cmd = tolower(ProcessCommandLine)
| where lower_cmd matches regex ".*mimikatz.*"

// OPTIMIZED - early filtering, uses has
DeviceProcessEvents
| where Timestamp > ago(24h)
| where ProcessCommandLine has "mimikatz"
| project Timestamp, DeviceName, FileName, ProcessCommandLine

Materialize for Reuse

let suspiciousDevices = materialize(
    DeviceProcessEvents
    | where Timestamp > ago(1h)
    | where FileName =~ "powershell.exe"
    | where ProcessCommandLine has "-enc"
    | distinct DeviceId
);
DeviceFileEvents
| where DeviceId in (suspiciousDevices)
| where Timestamp > ago(1h)

General Tips

• Use let statements for readability and reuse
• Test queries with | take 100 before running full scope
• Use render for visualization: | render timechart
• Check schema with TableName | getschema
• Use datatable for inline reference lists
• Combine MITRE ATT&CK technique IDs in comments for documentation