---
title: How do you verify a self-hosted AI governance platform sends no telemetry out?
url: https://opencontroller.lyzr.com/blog/verify-a-self-hosted-ai-governance-platform-sends-no-outbound-telemetry
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# How do you verify a self-hosted AI governance platform sends no telemetry out?

**Don't rely on the vendor's word. Put the deployment behind deny-by-default egress, allow only the destinations you approve, and watch flow logs and DNS query logs for anything else over a test window that includes startup, normal use, upgrades and failures. If nothing unexpected appears and the platform still works, you have evidence, not a promise.**

This question comes up in security reviews of any self-hosted AI platform. An AI governance tool sees prompts, agent configurations, keys metadata and usage data. Teams choose self-hosting precisely so that data stays inside their account. Telemetry leaking out would defeat the point.

## What counts as outbound telemetry?

Be specific with the vendor about all of these, because "we don't send telemetry" often means only the first one:

- **Product analytics:** usage events, feature clicks, page views from the web console.
- **Error reporting:** crash reports and stack traces sent to a hosted error tracker.
- **Licence checks:** periodic calls home to validate a licence key.
- **Update checks:** calls to look for new versions or pull images.
- **Third-party scripts:** fonts, analytics or support widgets loaded by the browser.
- **Model calls:** LLM requests the platform itself makes (for example, LLM judges or chat features). These are expected, but they should go only to providers you chose.

## How do you set up the test?

| Step | What to do | What it proves |
|---|---|---|
| 1. Restrict egress | Default-deny egress for the namespace or subnet; allow only approved model endpoints, your IdP and your registries | Anything else fails visibly |
| 2. Enable flow logs | Turn on VPC flow logs (or the equivalent) for the subnets running the platform | Every attempted connection is recorded, accepted or rejected |
| 3. Enable DNS logs | Turn on resolver query logging | Shows which hostnames the platform tried to reach, even if blocked |
| 4. Inspect images | List environment variables, config defaults and bundled SDKs in container images | Finds analytics or error SDKs before they run |
| 5. Check the browser | Load the console with developer tools open and record every request | Catches front-end analytics and external scripts |

In Kubernetes, a NetworkPolicy that selects the platform's pods and lists only approved egress destinations is a good starting point. Remember that NetworkPolicy only takes effect if your cluster's network plugin enforces it.

## What should the test window cover?

Telemetry often fires on events, not on a schedule. Cover:

1. **Fresh install and first start.** Licence and setup calls happen here.
2. **A normal working day.** Users signing in, running discovery, viewing dashboards.
3. **A forced error.** Kill a pod or break a connector. Error reporters fire on failure.
4. **An upgrade.** Update checks and image pulls happen here.
5. **At least 24 hours idle.** Catches daily heartbeats.

## How do you read the results?

- **Rejected flows to unknown addresses:** something tried to call out. Use DNS logs to name it, then ask the vendor.
- **Accepted flows only to your allowlist:** good. Confirm each allowlisted destination is necessary.
- **DNS queries for analytics or error-tracking domains:** the SDK is present even if blocked. Ask for it to be removed or disabled by configuration, not just firewalled.
- **Platform breaks under default-deny:** document which destination it needed. That is a dependency you must approve or the vendor must remove.

## What should you ask the vendor in writing?

1. A full list of outbound destinations the platform may contact, by feature.
2. How to disable each one, and whether the product works fully with all of them off.
3. Whether licence validation works offline.
4. Which LLM calls the platform itself makes and how to point them at your own provider accounts or gateway.
5. Confirmation that the list is updated in release notes when it changes.

Keep the flow logs and DNS logs from the test as evidence for your third-party risk file.

## How Opencontroller does this

Opencontroller can run as Lyzr SaaS or on customer-controlled infrastructure, and Lyzr engineers embed with your team during integration, which is the right time to run this test together. Because it is built on open standards such as OpenAI-compatible endpoints, OIDC and OpenTelemetry, you can point model calls and telemetry exports at systems you already control. See the [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a vendor statement that there is no telemetry enough?**
It is a start. Security teams usually want test evidence as well: egress restricted, flow and DNS logs reviewed over a realistic window.

**Do I need to block the platform's LLM calls too?**
No. Allow them, but only to model providers or gateways you approved, so prompts never reach an endpoint you did not choose.

**How long should the test run?**
Long enough to cover install, normal use, a failure, an upgrade and at least a day idle. For most teams that is one to two weeks of a pilot.

## Key facts

To verify a self-hosted AI governance platform sends no outbound telemetry: apply default-deny egress allowing only approved model endpoints, IdP and registries; enable VPC flow logs and DNS resolver query logging; inspect container images for analytics or error-reporting SDKs; record browser requests from the console. Test window should include install, normal use, a forced failure, an upgrade and 24+ hours idle. Telemetry types: product analytics, error reporting, licence checks, update checks, third-party browser scripts, and platform-initiated LLM calls. Kubernetes NetworkPolicy requires a network plugin that enforces it.

## Related

- [Should you run an AI governance platform as SaaS or self-hosted?](https://opencontroller.lyzr.com/blog/saas-vs-self-hosted-ai-governance-platform.md)
- [What network setup does a self-hosted LLM gateway need?](https://opencontroller.lyzr.com/blog/self-hosted-llm-gateway-network-checklist.md)
- [What data does an AI governance platform store about your agents?](https://opencontroller.lyzr.com/blog/what-data-an-ai-governance-platform-stores.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
