Skip to content

SOC CrowdStrike Falcon: one issue per process tree, not per detection - #1250

Closed
scottbrumley wants to merge 1 commit into
mainfrom
soc-crowdstrike-falcon/suppress-by-process-tree
Closed

scottbrumley wants to merge 1 commit into
mainfrom
soc-crowdstrike-falcon/suppress-by-process-tree

Conversation

@scottbrumley

@scottbrumley scottbrumley commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

What this changes

SOC CrowdStrike Falcon - Endpoint All Alerts suppresses on aggregate_id (CrowdStrike's process-tree grouping) instead of composite_id (one detection), still for 1 hour. Rule, vendor contract (schemas/vendors/crowdstrike-falcon/falcon-detections.yaml), generated schema doc and both READMEs change together.

Why

A CrowdStrike attack raises one issue per Falcon detection: a Turla Carbon replay gives ~120 issues in one case within minutes. On deathstar about half of such a burst never gets its playbook started (runStatus: pending, playbook state: new), and some finished runs stay running. Reproduced three times on 11–12 Nov (70/123, 78/123, and 53/106 even when paced at 20 events per 3 min); the same pattern goes back to September. Fewer issues per attack keeps the queue clear. The platform side is being raised with Product separately.

Turla Carbon: 130 detections, 46 process trees.

Trade-off

The first detection in a tree raises the issue and sets its severity; later detections in the same tree within the hour raise no issue of their own. Their events stay in the dataset and in the case's investigation query.

Tested on deathstar

  • check_contribution passed (upload retried once after a tenant 500).
  • Suppression works across time: two detections in one fresh tree, 3 minutes apart → one issue; the second was suppressed.
  • Burst (unpaced): 123 issues from 45 trees — detections arriving together race the suppression, so an unpaced replay still floods. Real sensors and povkit's paced replays don't arrive that way.
  • Paced (20/3m, like povkit): see the comment below once it lands.

Replays need a run-unique aggregate_id now, or a rerun within the hour raises nothing: tools/replay_scenario.py in #1251 and povkit internal/replay in Palo-Cortex/povkit#33.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VK24hKuKXVJfuVHg4Min9y

Suppression keys on aggregate_id (CrowdStrike's process-tree grouping)
instead of composite_id (one detection), still for 1 hour. A Turla Carbon
replay drops from 130 issues to 46, all in the same case as before.

Bursts of ~120 issues into one case leave about half of them never started
by the playbook engine; fewer issues per attack keeps the queue clear. The
first detection in a tree sets that issue's severity; later detections in
the tree within the hour raise no issue of their own.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VK24hKuKXVJfuVHg4Min9y
@scottbrumley

Copy link
Copy Markdown
Contributor Author

Closing: suppression loses later detections in a tree (a Low first detection would hide a later High one from the automation rule) and folds away techniques the assessment needs. Replacing with an aggregating rule: one issue per process tree carrying the highest severity and every technique.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

version:patch Bug fix or hotfix → x.x.N

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant