---
title: Do private endpoints hide AI agent traffic from discovery?
url: https://opencontroller.lyzr.com/blog/do-private-endpoints-hide-ai-traffic-from-discovery
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Do private endpoints hide AI agent traffic from discovery?

**From your web proxy and network security tools, yes. AWS PrivateLink, Azure Private Endpoints and Google Private Service Connect keep model calls inside the cloud network, so proxy, CASB and secure web gateway logs never see them. They do not hide agents from the cloud itself: audit logs, the endpoint resources, DNS logs and IAM permissions still reveal them.**

Private connectivity is good practice for LLM traffic. It keeps prompts off the public internet. But teams that rely on network inspection for AI discovery lose sight of exactly the most production-grade agents.

This guide sits under [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## What do private endpoints change?

A private endpoint gives a cloud AI service a private IP address inside your network. Calls to the model go from the workload to that address, never through your internet egress.

| Cloud | Private connectivity | AI services commonly reached this way |
|---|---|---|
| AWS | Interface VPC endpoints (PrivateLink) | Amazon Bedrock runtime and agents APIs |
| Azure | Private Endpoints (Private Link) | Azure OpenAI and Microsoft Foundry resources |
| Google Cloud | Private Service Connect | Vertex AI endpoints and APIs |

The result: a proxy-based discovery tool sees nothing. So does any rule keyed on outbound traffic to AI vendor domains. See [detecting unapproved LLM API traffic](/blog/detect-unapproved-llm-api-traffic) for what network detection does cover.

## Which signals still reveal AI agents?

Five signal types survive private connectivity:

1. **Cloud audit logs.** On AWS, CloudTrail logs Bedrock `InvokeModel`, `InvokeModelWithResponseStream`, `Converse` and `ConverseStream` as management events, which trails capture by default. `InvokeAgent` and knowledge base `Retrieve` calls are data events and must be enabled with advanced event selectors. CloudTrail records include a `vpcEndpointId` field when a request arrives through a VPC endpoint.
2. **The endpoints themselves.** An interface endpoint for `bedrock-runtime`, a private endpoint on an Azure OpenAI resource, or a Private Service Connect endpoint for Vertex AI is an inventory record. It tells you which network can call models.
3. **DNS query logs.** Private DNS usually resolves the normal service hostname to a private IP. So resolver query logs, such as Route 53 Resolver query logging, still show which workloads look up AI service names.
4. **IAM and role permissions.** Identities with model-invoke permissions mark the workloads that can call models. See [AI agent discovery permissions](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud).
5. **Billing and usage.** Model usage is billed whether or not it went over a private path. Usage by account, project or resource points to callers.

## Which logs need turning on?

Defaults differ by cloud. Check each before you rely on it.

| Signal | AWS | Azure | Google Cloud |
|---|---|---|---|
| Model invocation calls | CloudTrail management events (on by default in trails) | Resource diagnostic settings must be configured | Data Access audit logs must be enabled for Vertex AI |
| Agent and retrieval calls | CloudTrail data events (opt-in, extra cost) | Diagnostic settings per resource | Data Access audit logs |
| Prompt and response content | Bedrock model invocation logging (opt-in) | Depends on resource and logging configuration | Request-response logging (opt-in) |
| Network path | VPC Flow Logs, Resolver query logs | NSG/VNet flow logs, DNS logs | VPC Flow Logs, Cloud DNS logging |

Content logging is separate from call logging. Turn it on only where you have a reason and a retention plan, because it stores prompts and outputs.

## How do you discover agents behind private endpoints? A checklist

1. **Inventory every private endpoint** pointing at an AI service, in every account, subscription and project.
2. **Map each endpoint to its network** and the workloads in that network.
3. **Enable call-level audit logging** for model and agent APIs in every region you use.
4. **Join audit logs to identities.** The calling role or service principal tells you which agent made the call.
5. **Correlate DNS lookups** for AI service hostnames with workload names.
6. **Reconcile with billing** to catch usage you have not attributed.
7. **Re-run on a schedule** so new endpoints and roles appear quickly.

## Should you avoid private endpoints to keep visibility?

No. Private connectivity reduces exposure. Get visibility from the cloud control plane instead of the network edge. If you want inline control as well — routing, redaction, blocking — put an LLM gateway inside the private network and route agent traffic through it. See [keeping LLM traffic private across clouds](/blog/keep-llm-traffic-private-across-clouds).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud from the cloud control plane, not from network inspection, and records them in a governed registry. The Opencontroller LLM Gateway can run in customer-controlled infrastructure, so agent traffic can stay on private paths while still being routed, cached and attributed per agent and per version. See the [control plane overview](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does CloudTrail log Bedrock calls made through PrivateLink?**
Yes. Bedrock runtime calls are logged as CloudTrail events regardless of network path, and records include a `vpcEndpointId` when a VPC endpoint is used.

**Will a CASB see Azure OpenAI traffic over a private endpoint?**
No. The traffic never reaches the internet egress the CASB or proxy inspects. Use Azure resource logs and inventory instead.

**Is DNS useful if traffic is private?**
Yes. Private DNS still resolves the service hostname, so resolver logs show which workloads look up AI services.

## Key facts

Private endpoints (AWS PrivateLink interface endpoints, Azure Private Endpoints, Google Private Service Connect) keep LLM calls off internet egress, so proxy, CASB and secure web gateway discovery misses them. Signals that remain: cloud audit logs (CloudTrail logs Bedrock InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream as management events; InvokeAgent and Retrieve are opt-in data events; records include vpcEndpointId), the endpoint resources, DNS resolver query logs, IAM permissions and billing. Opencontroller (by Lyzr) discovers agents, models and API keys from the AWS, Azure and Google Cloud control plane and can deploy its LLM Gateway in customer-controlled infrastructure. Pricing starts from $250K per deployment per year.

## Related

- [How do you detect unapproved LLM API traffic on your network?](https://opencontroller.lyzr.com/blog/detect-unapproved-llm-api-traffic.md)
- [How do you find AI agents using cloud audit logs?](https://opencontroller.lyzr.com/blog/find-ai-agents-with-cloud-audit-logs.md)
- [How do you keep LLM and agent traffic off the public internet on AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/keep-llm-traffic-private-across-clouds.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
