Documentation

Deploy the agent and build your inventory | AI Agent Security

A workflow for AI Agent Security

AI coding assistants arrive on your fleet without a purchase order. A developer installs one, it connects to an MCP server, and a marketplace skill changes how it behaves, none of which shows up in a traditional asset inventory. This workflow shows you how to deploy the AI Agent Security agent and produce a first-pass inventory of every AI agent, MCP server, LLM endpoint, and agent skill on your endpoints.

Scope of this guide

This workflow covers deploying the agent and reviewing the five Visibility pages to establish what AI activity exists in your environment. It does not cover writing policies or investigating individual sessions.

Prerequisites

  • You have administrator access to the tenant.
  • You have a deployment method available: BeyondTrust EPM with the Package Manager, your own software distribution tool, or local administrator rights on a test machine.
  • Your tenant root certificate is available from Settings > Administration, on the Trust anchor tab.

Steps

Step 1: Deploy the AI Agent Security agent

  1. Choose your deployment path. Where you already run BeyondTrust EPM, the Package Manager installs the agent to the computer groups you select, with no deployment key to copy and no install command to run. For other paths, go to Settings > Deployment.
  2. Distribute your tenant root certificate where it is required. Windows agents install the certificate themselves when they activate; macOS endpoints need it distributed with your device management tool. Endpoints must trust it before the agent can inspect HTTPS traffic.
  3. Wait for endpoints to check in. On the EPM path, this takes up to about 2 hours by default.
  4. Confirm the endpoints appear on the Endpoints page with a recent last-seen time.

Step 2: Confirm discovery is on

  1. Go to Settings > Administration.
  2. Select the Fleet defaults tab.
  3. Confirm MCP discovery and Skill discovery are on. An inventory that looks clean may only mean discovery is off.
  4. Check Scan interval to see how often endpoints run their discovery scans.

Step 3: Review the AI agents found

  1. Go to Visibility > AI Agents.
  2. Review the Endpoints column for each agent to see how far it has spread.
  3. Check the Type and Vendor columns to tell coding agents from desktop apps.
  4. Compare the list against the assistants your organization has sanctioned. Note any agent nobody approved.

Step 4: Review the MCP servers found

  1. Go to Visibility > MCP Servers.
  2. Read the cards at the top: Servers discovered, Blocked, Monitored or approval required, and Enforcement gaps.
  3. Select the Tools column heading to sort by it, so the servers that expose the most tools are easy to find.
  4. Check the Access column. A server with no access control is allowed, and every tool call to it is permitted.
  5. Note any server your team did not know was in use.

Step 5: Review the LLM endpoints found

  1. Go to Visibility > LLM Endpoints.
  2. Review the Host column for model providers your organization has not approved.
  3. Check Devices seen to see how widely each host is used.

Step 6: Review the agent skills found

  1. Go to Visibility > Agent Skills.
  2. On the Overview tab, read Skills with critical findings and Third-party skills.
  3. Select the Skills tab and filter on Severity for the most severe values.
  4. Filter on Source for marketplace, which means the skill came from a plugin marketplace rather than being written locally.

Step 7: Confirm endpoint coverage

  1. Go to Visibility > Endpoints.
  2. Read the platform tiles at the top, which count your endpoints by operating system.
  3. Filter on Status for Stale or Offline to find endpoints that stopped reporting.
  4. Filter on Enforcement for Off to find endpoints where enforcement is not active.

Verify the results

Confirm that at least one AI agent, MCP server, LLM endpoint, or agent skill appears that your IT or security team did not already know about. Confirm that the platform totals on the Endpoints page match the number of machines you deployed to. A shortfall points to machines where the agent is not installed, not enrolled, or not reporting.

Next steps

  • Review the Risky agent behavior card on the Dashboard for high-risk actions already blocked or audited on the endpoints you discovered.
  • Separate human and agent privilege on a machine where an unapproved agent is running.
  • Assign an Owner to endpoints showing a dash, so you know who to contact about each machine.

Updated about 2 hours ago

©2003-2026 BeyondTrust Corporation. All Rights Reserved. Other trademarks identified on this page are owned by their respective owners. BeyondTrust is not a chartered bank or trust company, or depository institution. It is not authorized to accept deposits or trust accounts and is not licensed or regulated by any state or federal banking authority.