Hunt and detection queries for the tools and techniques I write up on newsletter.securityinbits.com. Each markdown file is one detection topic and carries the query in both stacks where relevant:
- KQL for Microsoft Defender for Endpoint (Advanced Hunting / custom detections)
- ES|QL for Elastic (Sysmon / ECS)
Keeping both stacks in one topic file is deliberate: same detection idea, two stacks, side by side, with the prose and tuning notes that explain it. The repo is topic-first, not language-first, on purpose.
| File | Topic |
|---|---|
anydesk.md |
AnyDesk RMM abuse: rename-proof identity, prevalence, --set-password, relay FQDNs, operator-IP failed P2P + the 7 SigmaHQ rules that fired |
discovery-tools.md |
SoftPerfect NetScan, Advanced IP/Port Scanner (process / network / file events) |
softperfect-netscan.md |
SoftPerfect NetScan, focused process detection |
ssh-tunnel.md |
Reverse SSH SOCKS tunnel (ssh -R) detection + the aggregation tell |
screenconnect.md |
ScreenConnect RMM abuse: download origin (FileOriginUrl), vendor-signed client (match identity not trust), relay classifier (cloud 443 vs self-hosted 8041), the e=Access unattended-access launch string + the 3 SigmaHQ rules that fired |
The HEADING names the real language; the FENCE uses the closest GitHub-highlightable identifier.
- KQL ->
```kustofence (GitHub's Kusto grammar), under a## KQL .../### KQL ...heading - ES|QL ->
```sqlfence (generic SQL, the closest available match), under a## ES|QL .../### ES|QL ...heading