12 log analysis

Log Analysis & SIEM Integration

Purpose

Enable Claude to assist with security log analysis across all major platforms. Claude directly parses and analyzes log samples provided by the user, builds SIEM queries for any platform, creates Sigma rules for portable detection, develops correlation rules, and identifies anomalous patterns in log data.

---

Activation Triggers

This skill activates when the user asks about:

• Parsing Windows Event Logs, Linux syslog, or application logs
• Building Splunk SPL, Elastic KQL/EQL, QRadar AQL, or Sentinel KQL queries
• Creating Sigma rules for platform-agnostic detection
• Detecting anomalies or attack patterns in log data
• Building SIEM correlation rules for complex attack scenarios
• Converting queries between SIEM platforms
• Log source health monitoring and gap analysis
• Detecting lateral movement, privilege escalation, or persistence in logs
• EVTX analysis or Windows audit log review

---

Prerequisites

pip install pandas pyyaml python-dateutil

Platform tools:

• Splunk — Splunk Web, SPL, and SOAR
• Elastic Stack — Kibana, KQL, EQL
• Microsoft Sentinel — KQL, Workbooks
• IBM QRadar — AQL, Rules
• Sigma — Platform-agnostic rule format
• python-evtx — Parse Windows .evtx files without Windows

---

Core Capabilities

1. Log Parsing & Analysis

When the user pastes logs or provides log files:

Claude directly reads and analyzes logs to extract security-relevant events.

Windows Event Log — Critical Event IDs:

Event ID	Log	Description
4624	Security	Successful logon — Logon Type 3 (network) is interesting
4625	Security	Failed logon — track source IP for brute force
4648	Security	Logon with explicit credentials (RunAs)
4688	Security	New process created — needs CommandLine auditing enabled
4698	Security	Scheduled task created
4702	Security	Scheduled task updated
4720	Security	User account created
4728/4732	Security	Member added to security/local group
4768/4769	Security	Kerberos TGT/TGS requested
4776	Security	NTLM authentication
4946	Security	Windows Firewall rule added
5140	Security	Network share accessed
5145	Security	Network share file access
7045	System	New service installed
1102	Security	Audit log cleared
4103/4104	PowerShell	PowerShell module/script block logging

Linux Log Analysis — Key Patterns:

# Failed SSH logins
grep "Failed password" /var/log/auth.log | awk '{print $1,$2,$3,$11}' | sort | uniq -c | sort -rn

# Successful logins after failures (brute force success)
grep "Accepted password\|Accepted publickey" /var/log/auth.log

# Sudo usage
grep "sudo:" /var/log/auth.log | grep -v "session"

# Cron job execution
grep CRON /var/log/syslog

# New user creation
grep "useradd\|usermod" /var/log/auth.log

# Privilege escalation
grep "su\b" /var/log/auth.log

Log parsing script:

python scripts/log_parser.py --input /var/log/auth.log --format json --output parsed.json
python scripts/log_parser.py --input events.evtx --normalize ecs --output normalized.json

2. SIEM Query Library

When the user asks to build detection queries:

Splunk SPL — Attack Pattern Queries

// Brute force attack detection
index=windows EventCode=4625
| bin _time span=5m
| stats count as FailedLogins, values(Account_Name) as Accounts by src_ip, _time
| where FailedLogins > 20
| sort -FailedLogins

// Pass-the-Hash detection (Logon Type 3 with NTLM)
index=windows EventCode=4624 Logon_Type=3 Authentication_Package=NTLM
| where NOT (Account_Name="ANONYMOUS LOGON" OR Account_Name="*$")
| stats count by Account_Name, Workstation_Name, src_ip
| where count > 1

// Lateral movement via PsExec / admin shares
index=windows EventCode=5145
| where (ShareName="\\\\*\\ADMIN$" OR ShareName="\\\\*\\C$") 
    AND RelativeTargetName="*PSEXESVC*"
| table _time, SubjectUserName, IpAddress, ShareName

// PowerShell encoded command execution
index=windows (source="WinEventLog:Microsoft-Windows-PowerShell/Operational" EventCode=4104)
    OR (EventCode=4688 CommandLine="*powershell*")
| search CommandLine IN ("*-EncodedCommand*", "*-enc *", "*-e *", "*-nop*", 
                          "*DownloadString*", "*IEX*", "*Invoke-Expression*")
| table _time, ComputerName, User, CommandLine

// Scheduled task creation for persistence
index=windows EventCode=4698
| rex field=TaskContent "<Command>(?P<command>[^<]+)</Command>"
| where NOT match(command, "(?i)\\\\windows\\\\|\\\\microsoft\\\\|\\\\system32\\\\")
| table _time, ComputerName, SubjectUserName, TaskName, command

// LSASS memory access (credential dumping)
index=sysmon EventCode=10 TargetImage="*lsass.exe"
| where NOT (SourceImage IN 
    ("C:\\Windows\\System32\\*", "C:\\Windows\\SysWOW64\\*",
     "C:\\Program Files\\*", "C:\\Program Files (x86)\\*"))
| table _time, SourceImage, GrantedAccess, CallTrace

// DCSync detection
index=windows EventCode=4662 
    (ObjectType="*domainDNS*" OR ObjectType="*19195a5b-6da0-11d0-afd3-00c04fd930c9*")
    (Properties="*Replicating Directory Changes All*" OR Properties="*1131f6ad*")
| where NOT match(SubjectUserName, "(?i)^.*\$$") 
| table _time, SubjectUserName, SubjectDomainName, Properties

// Kerberoasting detection
index=windows EventCode=4769 Ticket_Encryption_Type=0x17
| where NOT (Account_Name="*$" OR Service_Name IN ("krbtgt", "kadmin/changepw"))
| stats count by Account_Name, Client_Address, Service_Name
| where count > 3

Microsoft Sentinel KQL — Queries

// Impossible Travel (logins from geographically impossible locations)
let TimeDelta = 2h;
SigninLogs
| where ResultType == 0  // Successful logins only
| where TimeGenerated > ago(7d)
| project UserPrincipalName, Location, TimeGenerated, IPAddress
| sort by UserPrincipalName asc, TimeGenerated asc
| serialize
| extend PreviousLogin = prev(TimeGenerated), PreviousLocation = prev(Location)
| where UserPrincipalName == prev(UserPrincipalName)
| extend TimeDiff = TimeGenerated - PreviousLogin
| where TimeDiff < TimeDelta and Location != PreviousLocation
| project UserPrincipalName, Location, PreviousLocation, TimeDiff, IPAddress

// Azure AD privilege escalation
AuditLogs
| where OperationName in ("Add member to role", "Add eligible member to role")
| extend TargetUser = tostring(TargetResources[0].userPrincipalName)
| extend RoleAdded = tostring(TargetResources[0].displayName)  
| where RoleAdded in ("Global Administrator", "Security Administrator", 
                       "Exchange Administrator", "SharePoint Administrator")
| project TimeGenerated, TargetUser, RoleAdded, 
           InitiatedBy=tostring(InitiatedBy.user.userPrincipalName)

// Suspicious PowerShell activity
SecurityEvent
| where EventID == 4104
| where TimeGenerated > ago(24h)
| where ScriptBlockText has_any("IEX", "DownloadString", "EncodedCommand", 
                                  "WebClient", "Invoke-Expression", "bypass", "-nop")
| project TimeGenerated, Computer, Account, ScriptBlockText
| extend RiskScore = case(
    ScriptBlockText has "IEX" and ScriptBlockText has "DownloadString", 10,
    ScriptBlockText has "EncodedCommand", 7,
    ScriptBlockText has "bypass", 5, 3)
| where RiskScore >= 5
| order by RiskScore desc

Elastic EQL — Sequence Detection

// Detect fileless malware execution chain
sequence by host.name with maxspan=5m
  [process where event.type == "start" and
   process.name in ("outlook.exe", "winword.exe", "excel.exe")]
  [process where event.type == "start" and
   process.name in ("powershell.exe", "cmd.exe", "wscript.exe", "cscript.exe")]
  [network where network.direction == "egress" and
   not network.destination.ip in ("127.0.0.0/8", "10.0.0.0/8", "192.168.0.0/16")]

// Ransomware detection: mass file extension changes + shadow copy deletion
sequence by host.name with maxspan=30m
  [file where event.type == "creation" and
   file.extension in ("locked", "encrypted", "crypted", "enc", "readme")]
  [file where event.type == "creation" and
   file.name in ("README.txt", "DECRYPT.txt", "HOW_TO_DECRYPT.txt")]
  [process where event.type == "start" and
   process.command_line : "* delete shadows *"]

3. Anomaly Detection Methodology

When the user asks to detect anomalies in log data:

Statistical Anomaly Detection:

# Claude's approach to analyzing log data for anomalies:
import pandas as pd
from datetime import timedelta

# 1. Volume anomalies
# Calculate rolling average, flag if current > mean + 3*stddev

# 2. Time-based anomalies (off-hours activity)
# Business hours: Mon-Fri 08:00-18:00 local time
# Flag: admin activities on weekends, logins at 03:00 UTC

# 3. Never-before-seen entities
# - New admin account created
# - First-time login from country
# - New process never seen before
# - New domain in DNS queries

# 4. Impossible travel
# Calculate geographic distance / time delta
# Flag if impossible to travel physically in the time window

python scripts/anomaly_detector.py --logs parsed.json --baseline baseline.json --output anomalies.json

Anomaly Categories:

Category	Indicators
Volume spike	10x normal event rate in 5 minutes
Off-hours activity	Admin access at 03:00 local time
New geography	Login from country with no prior history
New process	First-ever execution of binary
Large data transfer	Upload > 10x baseline for this user/system
Silent log source	No events received in 30+ minutes
Authentication pattern	Logon Type 3 from non-admin workstation

4. Sigma Rule Development

When the user asks to create Sigma rules:

title: Credential Dumping via Procdump
id: e5eb5a27-4a98-4c34-8b39-1fbe552d2aa4
status: stable
description: Detects the use of ProcDump to dump LSASS memory for credential theft
author: SOC Analyst
date: 2025/05/28
references:
  - https://attack.mitre.org/techniques/T1003/001/
  - https://docs.microsoft.com/en-us/sysinternals/downloads/procdump
tags:
  - attack.credential_access
  - attack.t1003.001
logsource:
  category: process_creation
  product: windows
detection:
  selection_tool:
    Image|endswith:
      - '\procdump.exe'
      - '\procdump64.exe'
  selection_lsass:
    CommandLine|contains:
      - 'lsass'
      - '-ma 4'      # PID 4 = System, sometimes used
  selection_flags:
    CommandLine|contains|all:
      - '-accepteula'
      - '-ma'
  condition: selection_tool and (selection_lsass or selection_flags)
falsepositives:
  - Legitimate use by administrators for debugging (rare, should be investigated)
level: high

Sigma rule conversion to SIEM platforms:

# Install sigma-cli
pip install sigma-cli

# Convert to Splunk SPL
sigma convert -t splunk -p splunk_windows sigma_rule.yml

# Convert to Elastic KQL
sigma convert -t elasticsearch -p ecs_windows sigma_rule.yml

# Convert to Microsoft Sentinel KQL  
sigma convert -t kusto sigma_rule.yml

5. Correlation Rule Development

When the user asks to create correlation rules for multi-event detection:

## Correlation Rule: Brute Force → Successful Login → Lateral Movement

**Trigger:** 
  Event 1: 4625 (Failed Login) × 20+ in 5 minutes (same source IP)
  THEN Event 2: 4624 (Successful Login) from same source IP within 10 minutes
  THEN Event 3: 5145 (Admin Share Access) from same host within 30 minutes

**Logic:**

Step 1: Bucket failed logins by source IP in 5-minute windows Step 2: If count > 20 → mark IP as "brute force source"
Step 3: Watch for successful login from same IP within 10 minutes Step 4: If successful login → escalate to HIGH severity Step 5: Watch for lateral movement from the successfully logged-in host Step 6: Declare incident if all 3 events observed

**Splunk Correlation (corr_rule.conf):**
```spl
index=windows EventCode=4625
| bin _time span=5m
| stats count by src_ip, _time
| where count > 20
| join src_ip [
    search index=windows EventCode=4624 
    | bin _time span=10m
    | stats count by src_ip, _time, Account_Name, Workstation_Name
]
| table _time, src_ip, Account_Name, Workstation_Name, count

Suppression & De-duplication:

• Suppress same correlation alert from same source IP for 1 hour after first fire
• Exclude known vulnerability scanners (add scanner IP ranges to exception list)
• Exclude service accounts with documented scheduled tasks

---

Log Source Health Monitoring

When the user asks about log source health:

// Splunk: Detect silent log sources (no events in 30 minutes)
| tstats count WHERE index=* BY host, sourcetype, _time span=30m
| where _time > relative_time(now(), "-30m@m")
| stats max(_time) as last_seen by host, sourcetype
| where last_seen < relative_time(now(), "-30m@m")
| eval lag = round((now() - last_seen) / 60, 1)
| table host, sourcetype, last_seen, lag
| sort -lag

---

Script Reference

log_parser.py

python scripts/log_parser.py --input /var/log/auth.log --format json --output parsed.json
python scripts/log_parser.py --input events.evtx --normalize ecs --output normalized.json

anomaly_detector.py

python scripts/anomaly_detector.py --logs parsed.json --baseline baseline.json --output anomalies.json

---

Skill Integration

Condition	Adjacent Skill
Anomaly found → escalate to SOC	→ Skill 11 (CSOC Automation)
Anomaly is a hunt lead	→ Skill 06 (Threat Hunting)
Build timeline from logs for IR	→ Skill 07 (Incident Response)
Create detection rules from findings	→ Skill 15 (Blue Team Defense)

---

References

• Splunk SPL Reference
• Elastic EQL Documentation
• Microsoft Sentinel KQL
• Sigma Rules HQ Repository
• Windows Security Audit Policy
• OCSF Schema
• SANS Windows Logging Cheat Sheet

---

v3.0 Enhancements (2026 Update)

Normalized, testable detection content:

• OCSF / ECS normalization — map sources to the Open Cybersecurity Schema Framework (or Elastic Common Schema) so one detection works across feeds; state the schema in each rule.
• Sigma correlation rules — use Sigma's correlation extension (count/temporal/value-count) for multi-event detections (e.g., brute force → success, low-and-slow exfil), not just single-event matches.
• Detection-as-code CI — every rule has unit tests with positive/negative sample events; rules are linted and converted per-backend (sigma convert) in CI before deployment.
• UEBA & identity analytics — baseline per-user/host behavior; alert on deviation (new admin action, abnormal data volume, impossible travel) rather than fixed thresholds.
• Platform currency — examples for Splunk (tstats/data models), Microsoft Sentinel (KQL, ASIM functions), and Elastic (ES|QL/EQL).

Precision rule: each rule documents data source + schema, false-positive conditions, ATT&CK mapping, and a tested sample event.