security-alert-triage

Alert Triage

Analyze Elastic Security alerts one at a time: gather context, classify, create a case, and acknowledge. This skill depends on the case-management skill for case creation.

Prerequisites

Install dependencies before first use from the skills/security directory:

cd skills/security && npm install

Set the required environment variables (or add them to a .env file in the workspace root):

export ELASTICSEARCH_URL="https://your-cluster.es.cloud.example.com:443"
export ELASTICSEARCH_API_KEY="your-api-key"

Quick start

All commands from workspace root. Always fetch → investigate → document → acknowledge. Call the tools directly — do not read the skill file or explore the workspace first.

Common multi-step workflows

Always complete the full workflow: fetch → investigate → document → acknowledge. Do not stop after gathering context — create or update a case with findings before acknowledging.

Critical execution rules:

• Start executing tools immediately — do not read SKILL.md, browse the workspace, or list files first.
• For ES|QL queries, write the query to a temporary .esql file then pass it via --query-file. Do not use edit_file — use a single shell call with echo "..." > query.esql && node ... --query-file query.esql.
• Keep context gathering focused: run 2-4 targeted queries (process tree, network, related alerts), not 10+.
• Report only what tools return. Copy identifiers verbatim — do not paraphrase IDs, timestamps, or hostnames.

Critical principles

• Do NOT classify prematurely. Gather ALL context before deciding benign/unknown/malicious.
• Most alerts are false positives, even if they look alarming. Rule names like "Malicious Behavior" or severity "critical" are NOT evidence.
• "Unknown" is acceptable and often correct when evidence is insufficient.
• MALICIOUS requires strong corroborating evidence: persistence + C2, credential theft, lateral movement — not only suspicious API calls.
• Report tool output verbatim. Copy IDs, hostnames, timestamps, and counts exactly as returned by tools. Do not round numbers, abbreviate IDs, or paraphrase error messages.

Workflow

When triaging multiple alerts, group first, then triage each group:

Step 0: Group alerts before triaging

When the user asks about multiple open alerts, group them first to avoid redundant investigation: query open alerts, group by agent.id, sub-group by time window (~5 min = likely one incident), triage each group as a single unit.

Use ES|QL for an overview (write to file first for PowerShell):

For full query templates, see references/classification-guide.md.

Step 1: Check existing cases

Before creating a new case, check if this alert belongs to an existing one. Use the case-management skill:

Look for cases with the same agent ID, user, or related detection rule within a similar time window.

Note: find --search may return 500 errors on Serverless. Use find --tags or list instead.

Step 2: Gather context

This is the most important step. Do not skip or shortcut it. Complete ALL substeps before forming any classification opinion.

Time range warning: Alerts may be days or weeks old. NEVER use relative time like NOW() - 1 HOUR. Extract the alert's @timestamp and build queries around that time with +/- 1 hour window.

Substeps: (2a) Related alerts on same agent/user; (2b) Rule frequency across env (high = FP-prone); (2c) Entity context — process tree, network, registry, files; (2d) Behavior investigation — persistence, C2, lateral movement, credential access.

Example — process tree (use ES|QL with KEEP; avoid --full which produces 10K+ lines):

For full query templates and classification criteria, see references/classification-guide.md.

Step 3: Create or update case

After gathering context, create a case and attach alert(s). Use --rule-id and --rule-name (required; 400 error without them):

Case description: Summary (1-2 sentences); Attack chain; IOCs (hashes, IPs, paths); MITRE techniques; Behavioral findings; Response context (remediation, credentials at risk).

Step 4: Acknowledge alerts

Acknowledge ALL related alerts together. Use --dry-run first to confirm scope, then run without it:

Increase --window for longer attack chains (e.g., 300 for 5 minutes). Report the exact count of acknowledged alerts from the tool output. Pass --yes to skip the confirmation prompt (required when called by an agent).

Step 5: Repeat

Tool reference

fetch-next-alert.js

Fetches the oldest unacknowledged Elastic Security alert.

run-query.js

Runs KQL or ES|QL queries against Elasticsearch.

PowerShell warning: ES|QL queries contain pipe characters (|) which PowerShell interprets as shell pipes. ALWAYS use --query-file for ES|QL:

KQL queries without pipes can be passed directly:

acknowledge-alert.js

Acknowledges alerts by updating workflow_status to acknowledged.

Examples

• "Fetch the next unacknowledged alert and triage it"
• "Investigate alert ID abc-123 — gather context, classify, and create a case if malicious"
• "Process the top 5 critical alerts from the last 24 hours"

Guidelines

• Report only tool output — do not invent IDs, hostnames, IPs, or details not present in the tool response.
• Preserve identifiers from the request — use exact values the user provides in tool calls and responses.
• Confirm actions concisely using the tool's return data.
• Distinguish facts from inference — label conclusions beyond tool output as your assessment.

Production use

• All write operations (acknowledge-alert.js) prompt for confirmation. Pass --yes or -y to skip when called by an agent.
• Use --dry-run before bulk acknowledgments to preview scope without modifying data.
• The acknowledge script uses the Kibana Detection Engine API, which is compatible with both self-managed and Serverless deployments.
• Verify environment variables point to the intended cluster before running any script — no undo for acknowledgments.

Environment variables