> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.fabricgateway.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.fabricgateway.ai/_mcp/server.

# Review security and data handling

> See where traffic and secrets go, what Postman can access, and how the gateway behaves when something fails.

A security review of an AI gateway needs a few facts: which party sees your traffic, where your upstream secrets live, what leaves your network, and what happens when a component fails. This topic collects those facts, which the rest of the docs describe in more detail. It describes how the product behaves today and isn't a statement of compliance or certification.

## Know who sees your traffic

|                                      | Free (hosted)                                                                                                                                                                            | Enterprise (hybrid)                                       |
| ------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- |
| **Where requests are processed**     | The data plane in Postman cloud                                                                                                                                                          | The data plane in your VPC                                |
| **Where requests go next**           | From Postman cloud to your upstreams                                                                                                                                                     | From your network straight to your upstreams              |
| **Postman access to the data plane** | Postman runs it and shares it across organizations. Each organization's configuration and credentials are kept separate. Postman's infrastructure processes your requests and responses. | None. Postman has zero inbound access to your data plane. |

On the Enterprise (hybrid) deployment, Postman can't connect to your data plane. On the hosted plan, Postman operates the data plane, as the table above shows. Your data plane applies the configuration that the control plane sends, so anyone who can change your organization's configuration, in the console or through the Admin API, changes what your data plane does. The control plane can't run commands on your data plane.

For the full comparison, see [Choose a deployment model](/evaluate/deployment-models).

## Know where secrets live

By default, callers don't hold upstream credentials. The gateway removes the caller's credential and adds the upstream credential itself, so agents don't have direct access to provider keys. The exception is a route where you let callers send their own provider keys, such as one that uses the [Passthrough plugin](/reference/plugins/passthrough). On Passthrough routes, on Lightyear routes with **Use the client's provider credentials** on, and on REST routes with no upstream credential configured, the caller sends its own provider credential.

* **Free:** Postman stores your secrets in a Postman-managed vault. When you enter a secret in the console, your browser encrypts it, and the control plane decrypts it and passes it to the hosted data plane's vault.
* **Enterprise:** only references to credentials sync to the control plane. The secret values stay in your own vault.

For the four kinds of credentials and where each one goes, see [Understand credentials and keys](/concepts/credentials-and-keys).

## Know what leaves your network

On the Enterprise plan, the data plane sends the control plane configuration acknowledgments, heartbeats, usage counts, a metering record for each LLM request, health reports, and budget totals. Metering records include the consumer, provider, model, and token counts, but not prompts or responses. Traces, metrics, and logs export through OpenTelemetry to the backend that you configure. The console's **Traces** and **Analytics** pages show your traffic only when that backend is Postman's. For details, see [What leaves your network](/evaluate/deployment-models#what-leaves-your-network).

## Know how the gateway fails

* **Control plane outage:** the control plane never sits in the request path. The data plane keeps serving traffic with its last configuration, and a restarted container reloads it from an on-disk cache encrypted with your data plane's key. A new pod has no saved configuration until it reaches the control plane. With the distributed store, a replacement pod that reuses its persistent volume can restart from its saved data, but a newly added replica still has to enroll, so it can't start during an outage. For details, see [When the control plane is unreachable](/evaluate/deployment-models#when-the-control-plane-is-unreachable).
* **Plugin failure:** the built-in plugins that cover authentication, authorization, and redaction fail closed by default.
* **Network direction:** in an enterprise deployment, the data plane always dials out to the control plane over mutual TLS, and never accepts inbound network connections from it.

## Know what the gateway enforces

Access control works at the organization and namespace level, and plugins and grants can target a route, a connection, a consumer, or a consumer group. Plugins on each route can enforce authentication, rate limits, and PII redaction, every request leaves an access log record, and requests are traced when they're sampled. Usage is metered only when an upstream answered. For the plugins, see the [plugin reference](/reference/plugins). To follow one request through each check, see [Follow a request through the gateway](/concepts/life-of-a-request).

## Next steps

* [Choose a deployment model](/evaluate/deployment-models)
* [Know the current limitations](/evaluate/limits)