Skip to main content
Connect your AI

Use your security product with the endpoint agent

You may already run a security product that blocks AI apps or AI sites, or that checks what staff paste into them. StackJack can read that product's AI events. When it does, you can let the product do…

Written By Christopher Scaminaci

Last updated 13 days ago

You may already run a security product that blocks AI apps or AI sites, or that checks what staff paste into them. StackJack can read that product's AI events. When it does, you can let the product do a job and the endpoint agent steps back from that job. This page explains the choices, what StackJack needs from each product, and how to see who is doing each job right now.

At a glance

  • What it does: reads your security product's AI events and lets the product take over a job from the endpoint agent.
  • Plan: reading the events works on every plan. The jobs the agent can hand over need the Pro plan or higher.
  • Who: an Administrator.
  • Where: sources on AI Governance → Connected accounts → All sources. The Use my security product setting is on Staff computers → Endpoint agent.

Agent, my security product, automatic

The Endpoint agent page has a Use my security product setting for three jobs: the paste check, AI app control and AI site control. Each job has three choices.

  • The agent. The endpoint agent does the job on each machine, whatever else you run. Use it when you want the agent to be the one that enforces the job.
  • My security product. The agent stops doing the job and records that you chose this. Use it only when you know the product does the job on every machine.
  • Automatic. StackJack decides. A job goes to a product only when two things are true: the product is connected, and StackJack is reading that product's events. If either stops being true, the job goes back to the agent without you doing anything.

Under each job the page shows who has it right now. It says either Handed to the product, because StackJack reads its events, or Kept by the agent, with the reason. The reasons are: the source is off, this kind is paused, nothing has been read yet, the connection needs attention, nothing has been read for six hours, or the part of the product's feed that carries this job is not reading. A line appears only for a job that a product you connected could carry.

Handing a job to a product does not change your bill, and the product's events do not count a person toward your staff count. StackJack records the events as inventory facts.

What to set up

For each product, add a source on the All sources page with Add from a connection. Choose the product's report under What to read, choose the connection you made on the Connectors page, and give the source a name. The first read of a new source goes back seven days (30 days for Defender alerts, SentinelOne threats, CrowdStrike alerts, Huntress incident reports and GravityZone incidents, and 14 days for Huntress signals). A source reads nothing until it exists, so a product you connected but have not added as a source stays with the agent.

Each product should also allow the endpoint agent's signing certificate, so the product does not block the agent itself.

Microsoft Defender and Purview

Two sources read Microsoft Defender. Microsoft Defender alerts records detections and data loss prevention alerts about AI apps and sites. It records them as inventory facts and does not hand a job to Defender, because an alert does not prove that an app or a site was blocked. Microsoft Defender advanced hunting reads the AI sites reached, the blocks, the AI apps that ran, and the Purview data loss prevention events. It can carry all three jobs, one part of its feed at a time. If one part stops reading, only the job that part carries goes back to the agent.

What each needs:

  • The Microsoft Graph connection must be signed in by a person who holds the Security Reader role, or a Defender role that includes View data. The sources read with that person's rights.
  • If you use your own app registration, it needs the SecurityAlert.Read.All permission for the alerts source and the ThreatHunting.Read.All permission for the advanced hunting source. A missing permission stops only the source it belongs to.
  • For data loss prevention alerts, the built-in alert tuning rule that Microsoft introduced in October 2026 must be off. While it is on, Microsoft records those signals as behaviors, not alerts, and the alerts source cannot see them.
  • For data loss prevention events in advanced hunting, your Microsoft tenant must share Insider Risk data with Defender. Until it does, that table is empty. StackJack cannot check either setting for you, so a source with them off looks healthy and finds no data loss prevention events.

When you add a source for a client organization, name the client's Microsoft directory (its tenant id or a verified domain name). Your own organization leaves it empty.

DNSFilter

DNSFilter DNS query logs reads the AI traffic that DNSFilter saw and carries the AI site control job.

  • Create an API token in DNSFilter and add it as the DNSFilter connection.
  • Enter the DNSFilter organization id when you add the source. One DNSFilter key can reach many organizations, and a source reads only the one you name, so the id is always required.

Cisco Umbrella

Cisco Umbrella DNS activity reads the DNS lookups that Umbrella put in its generative AI category and carries the AI site control job.

  • Create the Umbrella API key with the reports.granularEvents:read and reports.utilities:read scopes.
  • Use one key for each organization. The source reads the organization that owns the key, so it has no organization id box. For a client organization, make a separate Umbrella connection that holds that client's own key.

ThreatLocker

ThreatLocker app-control blocks reads the executions that ThreatLocker denied and carries the AI app control job. It reads denials only, not the executions it allowed.

  • Create an API user in ThreatLocker that can read the Unified Audit, and add its token as the ThreatLocker connection.
  • For a client organization, enter the ThreatLocker child organization id when you add the source. Your own organization leaves it empty.

SentinelOne

SentinelOne threats reads the threats SentinelOne found and carries the AI app control job. A threat counts for the job when the blocklist or an application-control rule found it and SentinelOne mitigated it. SentinelOne STAR alerts records the alerts of your custom rules about AI apps and sites. A rule match is not a block, so that source records and reports and does not hand a job to SentinelOne.

  • Create a service user in SentinelOne whose role can read threats and alerts, and add its API token as the SentinelOne connection.
  • Enter the SentinelOne account or site when you add each source: write account: and the account id, or site: and the site id. One SentinelOne console token reaches every account and site you manage, and a source reads only the one you name, so this is always required.

CrowdStrike Falcon

CrowdStrike Falcon alerts records the Falcon alerts about AI apps and sites. StackJack reads and reports them, and the endpoint agent keeps the AI jobs.

  • Create an API client in Falcon with Alerts: Read, and add it as the CrowdStrike connection with the cloud region of your Falcon console.
  • Use one connection for each Falcon customer. The source reads the customer that owns the client, so it has no id box. For a client organization, make a separate CrowdStrike connection for that client's own customer ID.

Huntress

Huntress AI misuse incident reports reads the incident reports that Huntress itself classed as AI misuse. Huntress app control and process signals reads the app control and process signals about AI apps. Both are recorded and reported. Neither hands a job to Huntress.

  • Create an API key and secret key pair in Huntress, and add it as the Huntress connection.
  • Enter the Huntress organization id when you add each source. A Huntress account holds every client organization, and a source reads only the one you name, so the id is always required.

Bitdefender GravityZone

Bitdefender GravityZone incidents reads the incidents whose attack graph holds an AI app or an AI site. An incident is a whole graph, so StackJack records and reports it and never says that the app or the site was blocked. GravityZone sends its application control and content control events only by push, which StackJack does not read, so this source hands no job.

  • Create a GravityZone API key with the Incidents group, and add it as the GravityZone connection. The connection test also needs the General group.
  • For a client organization, enter the GravityZone company id (24 characters) when you add the source. Your own organization leaves it empty. The source reads that one company and never its child companies.

ESET PROTECT

ESET PROTECT detections reads the detections that ESET Connect recorded for AI apps and AI sites. It records and reports them and hands no job to ESET.

  • Create an ESET Connect API user that can read detections, and add it as the ESET connection with your ESET region.
  • One API user is one ESET account, so the source has no id box. For a client organization, make a separate ESET connection for that client.

Which products take a job

Of the products above, only SentinelOne threats hands a job to its product: the AI app control job. Every other source in this section is read and reported, and the endpoint agent keeps the job. A source that records detections still shows in your inventory and your evidence.

Check that it works

After you add a source, give it up to an hour for the first read. The All sources page shows when each source last delivered events. Then open the Endpoint agent page and look at the line under each job. If a line says Kept by the agent and gives the reason "the connection needs attention", open the Connectors page, check the connection, and test it again.

If a source cannot read for a short time, nothing changes: StackJack keeps the job with the product for up to six hours. If the credential is rejected, the job goes back to the agent at once. If nothing is read for six hours, it goes back then. In both cases the Endpoint agent page says so.