gensecaihq/wazuh-autopilot/wazuh-active-response

Wazuh Active Response

Agents never execute actions. The response-planner proposes (propose_action); the autonomy policy and human approvers decide; the platform executor calls the Wazuh MCP action tool; the responder verifies. This skill explains what happens underneath so plans are realistic and verification is honest.

How Wazuh AR works

1. The Wazuh API (via the MCP server) sends an active-response command to one or more agents.
2. The agent runs a script of that name from its active-response bin directory. If the script

isn't deployed on that agent, nothing happens on the host.

1. The API returns as soon as the manager accepts and dispatches the command

("AR command was sent to the agent") — this is not proof of execution.

1. The script logs its outcome; Wazuh turns that log into an alert in the active_response group.
2. API-triggered AR is permanent: Wazuh ignores the timeout for API-triggered commands, so a block

stays until explicitly removed (rollback), unlike rule-triggered AR with a timeout.

What each platform action sends

If the org hasn't deployed the custom scripts, isolate/kill/quarantine will be "dispatched" and do nothing. Say so in the plan and prefer actions with stock scripts, or recommend a manual step.

Blast radius

• block_ip with no agent means all_agents=true: fleet-wide and permanent. Name the agents that

actually see the attacker (from the alerts) instead, unless the IP is confirmed malicious and internet-only. Fleet-wide blocks can only be rolled back agent by agent.

• Never target the manager (agent 000) — the MCP server refuses it by default.
• Protected targets (policy) are refused before execution; don't propose them.
• Isolation cuts the host off except for the manager: confirm the business owner impact first.

Pre-checks (responder and response-planner)

1. Agent connected? check_agent_health agent_id="<id>" — AR to a disconnected agent is queued or lost.
2. Already done? wazuh_check_blocked_ip / wazuh_check_agent_isolation / wazuh_check_user_status

before proposing a duplicate.

1. Target still valid? For kill_process, confirm the PID with get_agent_processes (syscollector is a

periodic inventory — PIDs can be stale or reused).

1. Evidence preserved? Kill and quarantine destroy volatile evidence — record process details first.

Verification caveats

The wazuh_check_* tools infer state; they are not live host queries:

Unblock evidence: rule 602 / rule 652 (level 3, "Host Unblocked by firewall-drop Active Response"). rule 658 (level 5) means the AR was activated but may not have had an effect — treat as not verified. No AR alert at all usually means the script didn't run (missing script, agent offline, AR disabled).

Report verification as verified / not verified / inconclusive with the evidence, never as "done" because the API said so.

Roles

• response-planner: pick the least destructive action that contains the threat, order it

(evidence → contain → eradicate), include rollback and blast radius in the rationale, then propose. Hand off to responder only via the platform (verification runs automatically after execution).

• responder: run the matching wazuh_check_*, look for the AR alerts above, record the finding,

and if something went wrong recommend rollback — hand off to response-planner for any new action.

Output

add_finding with action, target agent(s), AR command, stock/custom script status, pre-check results, verification evidence (rule and timestamp) and rollback path; refs such as MITRE-D3FEND:D3-ITF, NIST-800-61r3:Respond.