security-detection-rule-management（安全检测规则管理）

Detection Rule Management

Create new detection rules for emerging threats and coverage gaps, and tune existing rules to reduce false positives. All operations use the Kibana Detection Engine API via rule-manager.js.

Execution rules

• Start executing tools immediately — do not read SKILL.md, browse the workspace, or list files first.
• Report tool output faithfully. Copy rule IDs, names, alert counts, exception IDs, and error messages exactly as returned by the API. Do not abbreviate rule UUIDs, invent rule names, or round alert counts.
• When a tool returns an error (rule not found, API failure), report the exact error — do not guess at alternatives.

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"
export KIBANA_URL="https://your-cluster.kb.cloud.example.com:443"
export KIBANA_API_KEY="your-kibana-api-key"

Common multi-step workflows

For endpoint behavior rules, always fetch the rule definition first to understand query logic and existing exclusions before adding an exception. For SIEM rules, always investigate alert patterns with run_query before tuning.

Critical: For endpoint behavior rules, always use fetch_endpoint_rule (not shell or direct script calls) to get the rule definition, then use add_endpoint_exception to add the exception. These are dedicated tools — do not invoke the underlying scripts manually.

Workflow: Tune a rule for false positives

Steps 1–2: Identify noisy rules and analyze false positives

Find noisy rules with noisy-rules or find, then get the rule definition and investigate alerts:

node skills/security/detection-rule-management/scripts/rule-manager.js noisy-rules --days 7 --top 20
node skills/security/detection-rule-management/scripts/rule-manager.js find --filter "alert.attributes.name:*Suspicious*" --brief
node skills/security/detection-rule-management/scripts/rule-manager.js get --id <rule_uuid>
node skills/security/alert-triage/scripts/run-query.js "kibana.alert.rule.name:\"<rule_name>\"" --index ".alerts-security.alerts-*" --days 7 --full

Look for patterns: same process/user/host → exception candidate; broad pattern → tighten query; legitimate software → exception; too broad → rewrite or adjust threshold.

Step 3: Choose a tuning strategy

In order of preference:

1. Add exception — Best for specific known-good processes, users, or hosts. Does not modify the rule query. Use when the rule is correct in general but fires on known-legitimate activity.
2. Tighten the query — Patch the rule's query to exclude the FP pattern. Best when the false positives stem from the query being too broad.
3. Adjust threshold / alert suppression — For threshold rules, increase the threshold value. For any rule type, enable alert suppression to reduce duplicate alerts on the same entity.
4. Reduce risk score / severity — Downgrade the rule's priority if it generates many low-value alerts but still has some detection value.
5. Disable the rule — Last resort. Only if the rule provides no value or is completely redundant with another rule.

Steps 4–5: Apply tuning, verify, and document

Add exception (single/multi-condition, wildcard via matches):

node skills/security/detection-rule-management/scripts/rule-manager.js add-exception \
  --rule-uuid <rule_uuid> \
  --entries "process.executable:is:C:\\Program Files\\SCCM\\CcmExec.exe" "process.parent.name:is:CcmExec.exe" \
  --name "Exclude SCCM" --comment "FP: SCCM deployment" --tags "tuning:fp" "source:soc" --yes

Patch query, threshold, severity, or disable:

node skills/security/detection-rule-management/scripts/rule-manager.js patch --id <rule_uuid> --query "process.name:powershell.exe AND NOT process.parent.name:CcmExec.exe" --yes
node skills/security/detection-rule-management/scripts/rule-manager.js patch --id <rule_uuid> --max-signals 50 --yes
node skills/security/detection-rule-management/scripts/rule-manager.js patch --id <rule_uuid> --severity low --risk-score 21 --yes
node skills/security/detection-rule-management/scripts/rule-manager.js disable --id <rule_uuid> --yes

Write operations (patch, enable, disable, delete, add-exception, bulk-action) prompt for confirmation by default. Pass --yes to skip the prompt (required when called by an agent).

Verify with rule-manager.js get --id <rule_uuid>. Update triage cases via the case-management skill.

Workflow: Create new detection rule

Steps 1–2: Define the threat, data sources, and fields

Specify MITRE ATT&CK technique(s), required data sources (Endpoint, Network, Cloud), and malicious vs legitimate behavior. Common indexes: logs-endpoint.events.process-*, logs-endpoint.events.network-*, .alerts-security.alerts-*, logs-windows.*, logs-aws.*. Key fields: process.name, process.command_line, process.parent.name, destination.ip, winlog.event_id, event.action. Verify data with run-query.js:

node skills/security/alert-triage/scripts/run-query.js "process.name:certutil.exe" --index "logs-endpoint.events.process-*" --days 30 --size 5

Step 3: Write and test the query

Rule types: query (KQL field matching), eql (event sequences), esql (aggregations), threshold (volume-based), threat_match (IOC correlation), new_terms (first-seen). Test against Elasticsearch before creating:

node skills/security/alert-triage/scripts/run-query.js "process.name:certutil.exe AND process.command_line:(*urlcache* OR *decode*)" \
  --index "logs-endpoint.events.process-*" --days 30

For EQL, use --query-file to avoid shell escaping issues.

Validate query syntax before creating or patching a rule. The validate-query command catches common errors locally — escaped backslashes, mismatched parentheses, unbalanced quotes, and duplicate boolean operators:

node skills/security/detection-rule-management/scripts/rule-manager.js validate-query \
  --query "process.name:taskkill.exe AND process.command_line:(*chrome.exe* OR *msedge.exe*)" --language kuery

The create and patch commands also run validation automatically and reject invalid queries. Pass --skip-validation only if you are certain the query is correct despite triggering a check.

Common KQL syntax mistakes:

• Escaped forward-slashes — KQL wildcards use plain text. Write */IM chrome.exe*, not *\/IM chrome.exe*.
• Mismatched parentheses — every ( must have a matching ).
• Unbalanced quotes — every " must be paired.
• Duplicate operators — AND AND or OR OR is always an error.

Step 4: Create the rule

node skills/security/detection-rule-management/scripts/rule-manager.js create \
  --name "Certutil URL Download or Decode" \
  --description "Detects certutil.exe used to download files or decode Base64 payloads, a common LOLBin technique." \
  --type query \
  --query "process.name:certutil.exe AND process.command_line:(*urlcache* OR *decode*)" \
  --index "logs-endpoint.events.process-*" \
  --severity medium --risk-score 47 \
  --tags "OS:Windows" "Tactic:Defense Evasion" "Tactic:Command and Control" \
  --false-positives "IT administrators using certutil for legitimate certificate operations" \
  --references "https://attack.mitre.org/techniques/T1140/" \
  --interval 5m --disabled

For complex rules (EQL sequences, MITRE mappings, alert suppression), use create --from-file rule_definition.json and --threat-file. See references/detection-api-reference.md for schema.

Step 5: Monitor and iterate

Monitor alert volume with noisy-rules --days 3 --top 10 and tune false positives as needed.

Workflow: Endpoint behavior rules tuning

Tune Elastic Endpoint behavior rules by adding Endpoint exceptions scoped to specific rules. Endpoint exceptions live in Security → Exceptions → Endpoint Security Exception List, not under individual SIEM rules.

Key principles: Always fetch the rule definition from protections-artifacts first. Always scope exceptions to the rule (rule.id or rule.name). Use full paths over process names. Run the mandatory entity cross-check (Step 4b) before any exception. Simulate impact (Step 5b) and aim for ≥60% noise reduction.

Scripts: fetch-endpoint-rule-from-github.js (get rule TOML by id), add-endpoint-exception.js (add to Endpoint Exception List; rule.id/rule.name required), check-exclusion-best-practices.js.

For the full step-by-step workflow (Steps 1–6), queries, and simulation templates, see references/endpoint-behavior-tuning-workflow.md. For exclusion best practices, see references/endpoint-rule-exclusion-best-practices.md.

Tool reference

rule-manager.js

All commands are run from the workspace root. All output is JSON unless noted.

Endpoint behavior tuning: fetch-endpoint-rule-from-github.js (get rule TOML by id), add-endpoint-exception.js (add to Endpoint Exception List; rule.id/rule.name required), check-exclusion-best-practices.js.

Exception entry format

Pass entries as field:operator:value. Operators: is, is_not, is_one_of, is_not_one_of, exists, does_not_exist, matches, does_not_match. Example: process.name:is:svchost.exe, file.path:matches:C:\\Program Files\\*.

Additional resources

• For full API schema details, see references/detection-api-reference.md
• For endpoint behavior tuning: references/endpoint-exceptions-guide.md, references/endpoint-rule-exclusion-best-practices.md
• For alert investigation during tuning, use the alert-triage skill
• For documenting tuning actions in cases, use the case-management skill

Examples

• "Find the noisiest detection rules from the last 7 days and help me tune one"
• "Add an exception to exclude SCCM from the suspicious PowerShell rule"
• "Create a new detection rule for certutil URL download or decode"

Guidelines

• Report only tool output. When summarizing results, quote or paraphrase only what the tools returned. Do not invent IDs, hostnames, IPs, scores, process trees, or other details not present in the tool response.
• Preserve identifiers from the request. If the user provides specific hostnames, agent IDs, case IDs, or other values, use those exact values in tool calls and responses — do not substitute different identifiers.
• Confirm actions concisely. After executing a tool, confirm what was done using the tool's return data. Do not fabricate internal IDs, metadata, or status details unless they appear in the tool response.
• Distinguish facts from inference. If you draw conclusions beyond what the tools returned (e.g., suggesting a MITRE technique based on observed behavior), clearly label those as your assessment rather than presenting them as tool output.
• Start executing tools immediately. Do not read SKILL.md, browse directories, or list files before acting.
• Report tool output verbatim. Copy rule IDs, names, alert counts, and error messages exactly as returned. Do not abbreviate UUIDs or round numbers.

Production use

• All write operations (create, patch, enable, disable, delete, add-exception, bulk-action, add-endpoint-exception) prompt for confirmation. Pass --yes or -y to skip when called by an agent.
• Endpoint exceptions suppress detections globally. Always scope exceptions to a specific rule using rule.id or rule.name in the entries. A broad, unscoped exception can silently reduce detection coverage.
• Verify environment variables point to the intended cluster before running any script.
• Use --dry-run with bulk-action to preview impact before executing bulk changes.

Environment variables