---
title: How do you find MCP servers exposed to the internet?
url: https://opencontroller.lyzr.com/blog/find-mcp-servers-exposed-to-the-internet
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you find MCP servers exposed to the internet?

**Scan your own external attack surface for MCP fingerprints: HTTP endpoints on paths such as `/mcp`, `/sse` or `/messages` that answer JSON-RPC `initialize` or `tools/list` requests. Combine that with an internal review of load balancers, ingress rules, developer tunnels and servers bound to all network interfaces. Any server that lists its tools without authentication should be taken offline or put behind authentication the same day.**

This is not a theoretical problem. In July 2025, researchers at Knostic used Shodan to identify 1,862 MCP servers reachable from the internet. They manually checked 119 of them, and every one of the 119 returned its internal tool list without authentication. Listing tools does not prove a tool can be called, but it does show that nothing stood in front of the server.

## Why do MCP servers end up exposed?

Several defaults and habits combine.

- **Authorization is optional in the MCP specification.** The spec defines an OAuth-based authorization flow for HTTP transports, but implementations are not required to use it.
- **Local servers bound to `0.0.0.0`.** The spec's transport guidance says local servers should bind only to localhost (127.0.0.1) and must validate the `Origin` header to prevent DNS rebinding. Example code and quick-start templates do not always follow that.
- **Developer tunnels.** Tools that publish a local port to a public URL for testing are easy to start and easy to forget.
- **Proof-of-concept deployments.** A team deploys a remote server to a cloud container service with a public load balancer to demo it, and it stays up.

## How do you scan for exposed MCP servers?

Only scan addresses and domains you own or are authorised to test.

1. **List your external surface.** Pull public IP ranges, DNS records, cloud load balancers and API gateways from your attack surface management tool or cloud inventories.
2. **Search internet scan data.** Search engines such as Shodan and Censys index HTTP responses. Query your own IP ranges and domains for MCP markers: JSON-RPC 2.0 bodies, `text/event-stream` responses, MCP-related paths and server headers from common Python and Node frameworks.
3. **Probe candidate endpoints.** Send an `initialize` request followed by `tools/list`. Do not call any tool. A response with a tool list and no authentication challenge is a finding.
4. **Check for an authentication challenge.** A protected server should return HTTP 401 with a `WWW-Authenticate` header pointing at its protected resource metadata.
5. **Review internal configuration.** Search ingress rules, security groups and container service configs for services named or labelled MCP with public exposure.
6. **Check developer tunnels.** Review egress logs and DNS for tunnelling services, and check device inventory for tunnel clients.

| Signal | Where to look | What it suggests |
|---|---|---|
| `tools/list` answered without credentials | Active probe | Unauthenticated server: highest priority |
| 401 with `WWW-Authenticate` | Active probe | Authentication in place; confirm token audience checks |
| Service listening on 0.0.0.0 on a laptop | Device inventory, local scan | Reachable from the local network |
| Public URL from a tunnelling service | DNS and egress logs | Temporary exposure that may persist |
| Public load balancer to an MCP container | Cloud config | Production or demo server exposed by design |

## What do you do once you find one?

Treat an unauthenticated MCP server as an incident, not a configuration note.

1. **Contain.** Remove public exposure or put the server behind authentication or an [MCP gateway](/blog/what-is-an-mcp-gateway).
2. **Rotate credentials the server holds.** MCP servers often carry API keys or database credentials. Assume they were reachable.
3. **Review logs.** Look for `tools/call` requests from unknown clients and for unusual downstream activity by the server's credentials.
4. **Find the owner.** Record the server in your MCP inventory with owner, purpose, tools and data reached. See [what to record for every MCP server](/blog/mcp-server-inventory-fields).
5. **Fix the pattern.** If a template or base image caused the exposure, fix it there, so the next team does not repeat it.

Once a server requires authentication, check that it validates token audience and does not pass client tokens through to upstream APIs. [How MCP authorization works](/blog/how-mcp-authorization-works) covers those requirements. For servers on developer machines, see [how to find local MCP servers on developer laptops](/blog/find-local-mcp-servers-on-developer-laptops). The pillar guide on [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) covers discovery across every source, and [how to govern MCP servers](/blog/how-to-govern-mcp-servers) puts these steps into a programme.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, tools and MCP servers, and Posture detects risky permissions and configurations. Opencontroller Edge extends discovery to local MCP servers on developer devices, installed by IT through MDM. External internet scanning is a job for your attack surface tooling; Opencontroller gives findings an owner and a policy. [Book a demo](https://www.lyzr.ai/book-demo/) to see it.

## Frequently asked questions

**Is it safe to send tools/list to a server we find?**
On infrastructure you own, a read-only `tools/list` is a low-impact check. Never call tools during discovery, and never probe systems you are not authorised to test.

**Is a stdio MCP server ever exposed to the internet?**
Not directly. A stdio server talks over a local pipe. It becomes reachable only if someone wraps it in an HTTP bridge or tunnel.

**Does an API gateway in front of the server fix the problem?**
Only if the gateway enforces authentication and authorisation per client. A gateway that just forwards traffic adds a hop, not a control.

## Key facts

Exposed MCP servers: Knostic research (July 2025) found 1,862 internet-reachable MCP servers via Shodan; all 119 manually checked returned tool lists without authentication. MCP authorization is optional in the spec; transport guidance says local servers should bind to 127.0.0.1 and must validate Origin to prevent DNS rebinding. Detection: own IP ranges and domains, Shodan/Censys queries for JSON-RPC and text/event-stream, probe initialize and tools/list only (never call tools), expect 401 with WWW-Authenticate when protected; check tunnels and 0.0.0.0 bindings. Response: contain, rotate held credentials, review tools/call logs, assign owner, fix template. Opencontroller (Lyzr) registry of MCP servers, Posture, Edge for local MCP servers. Pricing starts from $250K per deployment per year.

## Related

- [How does MCP authorization work?](https://opencontroller.lyzr.com/blog/how-mcp-authorization-works.md)
- [How do you find local MCP servers on developer laptops?](https://opencontroller.lyzr.com/blog/find-local-mcp-servers-on-developer-laptops.md)
- [Is a local stdio MCP server safer than a remote HTTP one?](https://opencontroller.lyzr.com/blog/mcp-stdio-vs-streamable-http-security.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
