Security ML Integrations

Security ML Integration Package Management

Install, check, or troubleshoot a security ML package. Supported packages: beaconing (Network Beaconing Identification), ded (Data Exfiltration Detection), dga (Domain Generation Algorithm Detection), lmd (Lateral Movement Detection), pad (Privileged Access Detection), problemchild (Living off the Land Attack Detection).

<!-- begin-partial: preamble -->

Environment Configuration

This skill executes Elasticsearch operations through the elastic CLI. If the elastic CLI is not installed, tell the user what it is needed for. Do not guess credentials, call the HTTP API directly, or attempt other workarounds.

This skill references operations in HTTP-shorthand form (e.g., GET /, GET /_cat/indices, GET /{index}/_mapping, GET /{index}/_settings/index.mode, POST /_query). The Operations table at the end of this document maps each shorthand to the equivalent elastic CLI command — always use the CLI rather than calling the HTTP API directly.

<!-- end-partial: preamble -->

Package source patterns, shared template fields, and ingest-pipeline wiring live in package capabilities.

Process

1. Classify the request. Status, README, or install. Status and README stay read-only — do not install, start transforms, deploy jobs, or enable rules. After identifying the package, branch: status uses GET kbn:/api/fleet/epm/packages/{package} only (title, version, status); README adds GET kbn:/api/fleet/epm/packages/{package}/{version}/docs/README.md and returns that document; install continues.

2. Identify the package. Map the user's words to one of the six package names above. If the name is ambiguous, ask before calling any API. Load that package's row from package capabilities — omit fields that are empty for that package.

3. Gather environment state before any mutation. Install only. GET /_license first — continue only for platinum, enterprise, trial, or Serverless. If GET /_ml/info returns 403, report missing manage_ml and stop. Then collect:

   • Package: GET kbn:/api/fleet/epm/packages/{package} for name, title, version, and status.

   • Source data: POST /{source_patterns}/_search with size: 0 and track_total_hits: true. If every pattern has zero hits, warn that jobs will not produce useful anomalies.

   • Shared template (DGA, PAD, ProblemChild only): GET /_component_template/{name}. Exists → plan a merge of the fields in the reference. 404 → plan a create with only those fields.

   • Shared pipeline (DGA, PAD, ProblemChild only): GET /_ingest/pipeline/{name}. Exists → plan to append a processor. 404 → plan a single-processor pipeline.

   • Transforms: GET /_transform/{pattern} (skip if the package has no transform pattern). Record each id, dest.index, and the .all alias (dest.aliases where move_on_creation is false).

   • Data views: GET kbn:/api/data_views and check whether the package's data-view pattern already exists.

   • ML capacity: GET /_ml/info (total memory) and GET /_ml/anomaly_detectors/_all/_stats (sum memory of opened jobs). GET kbn:/internal/ml/modules/get_module/{module_id} for job count and per-job memory. If jobs do not fit, say so and suggest prioritizing job groups — do not deploy anyway without the user accepting the risk.

   • Rules: GET kbn:/api/detection_engine/rules/prepackaged/_status and POST kbn:/internal/detection_engine/prebuilt_rules/installation/_review filtered by the package detection tag.

4. Present a plan and wait for confirmation. Show package identity and status, source-data hits (warn if all zero), template/pipeline merge-or-create, transform count and states, data-view presence, ML memory fit, and rule count. Do not call any write API until the user confirms. If they decline, stop.

5. Install the Fleet package, then apply shared mappings. After confirmation:

   • POST kbn:/api/fleet/epm/packages/{package} to install assets (transforms may be deferred until started).

   • If the package uses a component template: PUT /_component_template/{name} with the merged or new body. Then POST /{data_stream}/_rollover on the rollover target so new indices pick up mappings.

   • If the package uses an ingest pipeline: PUT /_ingest/pipeline/{name} with appended or new processors. Use the installed package version in the versioned pipeline name.

   Skip template, pipeline, and rollover when the package row has no values for them (beaconing, ded, lmd).

6. Start transforms and create destination aliases. Skip this block (except .ml-anomalies-shared) when the package has no transform pattern. Fleet's first checkpoint often skips .all aliases. After install, GET /_transform/{pattern} again if IDs were missing. Reset only if the transform is new, never started, failed, or the user asked to rebuild: POST /_transform/{id}/_stop, POST /_transform/{id}/_reset, POST /_transform/{id}/_update with the existing description, POST /_transform/{id}/_start, POST /_transform/{id}/_schedule_now. If already started with checkpointing.last.checkpoint >= 1, skip reset; restamp via the existing description and ensure .all aliases exist. Then POST /_aliases to add the .all alias (and .latest if the dest defines it) on dest.index. If a data-view pattern is missing, POST kbn:/api/data_views/data_view. Also ensure a hidden .ml-anomalies-shared data view exists (POST kbn:/api/data_views/data_view) so package dashboards resolve.

7. Confirm the first transform checkpoint before ML. If the package row has no transform pattern, skip this step and continue to jobs. Otherwise call GET /_transform/{pattern}/_stats. Proceed only when each transform is started and checkpointing.last.checkpoint >= 1; do not use documents_processed as a substitute. Zero documents with a completed checkpoint is acceptable when source patterns are empty — destination aliases still let datafeeds start. Otherwise tell the user transforms must finish processing and wait for confirmation before the next step. Never set up the ML module against missing dest indices.

8. Deploy ML jobs, then detection rules. After transforms are confirmed (or skipped):

   • POST kbn:/internal/ml/modules/setup/{module_id} with startDatafeed: true and indexPatternName set to the package data-view pattern, or source patterns when that cell is empty. Do not put spaces after commas in indexPatternName — the setup API splits on commas without trimming. Correct: logs-*,ml_okta_multiple_user_sessions_pad.all,ml_windows_privilege_type_pad.all.

   • Initialize the detection engine if needed (POST kbn:/api/detection_engine/index), install the prebuilt rules package (POST kbn:/api/fleet/epm/packages/security_detection_engine), review the package-tagged rules, then call POST kbn:/internal/detection_engine/prebuilt_rules/installation/_perform with {"mode":"SPECIFIC_RULES","rules":[{"rule_id":"<rule_id>","version":<version>}]} (one entry per reviewed rule). Enable the package's ML rules with POST kbn:/api/detection_engine/rules/_bulk_action (disable then enable on immutable ML rules tagged with the package detection tag so API keys are stamped with the calling user).

9. Report created versus skipped. Package version and asset count; template/pipeline/rollover created, updated, or skipped; transforms started; data views created or already present; ML jobs and datafeeds started (and any that failed, with the mapping or index reason); rules installed and enabled.

Examples

Install PAD

User: "Install PAD"

1. Gather state for pad. Package not installed, zero source hits, logs-endpoint.events.process@custom already has ProblemChild fields, no @custom pipeline, two stopped transforms, 15 jobs fit in available memory, tagged rules available.

2. Present the plan: warn about empty source data; merge process.command_line_entropy; create the PAD pipeline processor; start both transforms; create the PAD data view. Wait for confirm.

3. After confirm: POST kbn:/api/fleet/epm/packages/pad, merge-and-PUT the shared template, PUT the pipeline, restamp via existing description (reset only if transforms are new or stopped), POST /_aliases for each .all alias.

4. GET /_transform/*pad*/_stats. If checkpoint 1 is done (even with 0 docs), continue; otherwise wait for the user.

5. POST kbn:/internal/ml/modules/setup/pad-ml with a no-space indexPatternName, then install and enable rules tagged Use Case: Privileged Access Detection.

6. Report each section as created, updated, or skipped.

Install beaconing (no template or pipeline)

User: "Set up beaconing"

1. Gather state. Omit template, pipeline, and rollover — the beaconing row has none.

2. Present a shorter plan (package, one transform, data view, jobs, rules). Wait for confirm.

3. Install the package, start the transform, create ml_beaconing* if missing, confirm checkpoint, set up beaconing-ml, enable rules tagged Use Case: Network Beaconing Identification.

Status only

User: "Is beaconing installed?"

1. GET kbn:/api/fleet/epm/packages/beaconing only.

2. Report title, version, and status. A README request also fetches and returns the docs README. Do not install or start anything.

Guidelines

• Gather environment state before every install. Do not skip the pre-flight reads.

• Stay read-only on status and README questions.

• Wait for explicit confirmation before any write. Check license before the Fleet POST.

• Pause after transforms until checkpointing.last.checkpoint >= 1 (0 docs is acceptable when source patterns are empty). Skip transform and package data-view steps when those capability cells are empty.

• Never overwrite shared @custom templates or pipelines. PAD and ProblemChild share logs-endpoint.events.process@custom — merge fields and append processors.

• Put no spaces after commas in ML indexPatternName.

• Cite Fleet package APIs as kbn:/api/fleet/.... Do not use a fleet: prefix.

• Restamp a Fleet transform by setting --description to the existing text, then start. Do not use --body '{}'.

• On serverless, GET /_ml/info may return 410. Continue with opened-job stats from GET /_ml/anomaly_detectors/_all/_stats; do not abort the install.

Operations

Internal ML-module APIs, prebuilt-rules _review/_perform, detection-engine index, and rules disable/enable have no usable generated CLI commands yet. Keep the HTTP shorthand as the contract; do not add skill-owned scripts as a workaround. Generated Fleet EPM get/install commands exist — use those instead of the shorthand.