GHSA-w26r-fwg8-rcp3

### Summary MagicMirror applies `ipWhitelist` only as Express middleware, but the Socket.IO server is attached directly to the HTTP server without equivalent IP allowlist, origin, or namespace authentication checks. In a documented common deployment where MagicMirror listens on a non-loopback interface but expects `ipWhitelist` to restrict access, an untrusted network client can connect directly to module Socket.IO namespaces and send arbitrary module-helper notifications. This allows unauthenticated server-side requests through default modules and can reach command execution in the default `updatenotification` helper when a third-party module update is pending and the attacker supplies the update command through the trusted socket configuration path.

### Details The affected product is the npm package/application `magicmirror` at version `2.36.0`, tested at commit `fb41d24ef522e91e802e2a623ff6afbddeb3c9d8` from `https://github.com/MagicMirrorOrg/MagicMirror.git`.

Default committed settings bind to loopback and allow loopback only (`js/defaults.js:8-13`), so the remote network impact requires a documented common configuration where the server is reachable beyond loopback. The shipped sample explicitly documents non-loopback binding and IP allowlist behavior: `config/config.js.sample:11-20` says `address` may be another interface or `0.0.0.0`/`::`, and `ipWhitelist` controls allowed clients.

The trust-boundary issue is that Socket.IO is configured before and outside the Express middleware chain:

- `js/server.js:42-50` creates Socket.IO directly on the HTTP(S) server with `cors.origin: /.*$/`. - `js/server.js:89-90` applies `ipAccessControl(config.ipWhitelist)` only with `app.use(...)`, which protects Express routes and static files but not Socket.IO handshakes or namespaces. - A search of runtime files found no `allowRequest`, `io.use(...)`, handshake IP check, or namespace authentication for Socket.IO; the only relevant matches were `js/server.js:44` and `js/server.js:90`. - `js/node_helper.js:88-103` registers every module namespace and dispatches every socket event and payload directly to `socketNotificationReceived(...)`.

Once a client can reach the Socket.IO server, the following default-module server-side actions are reachable without an equivalent IP whitelist or module-authentication check:

- `defaultmodules/newsfeed/node_helper.js:12-17` accepts `CHECK_ARTICLE_URL` and calls `checkArticleUrl(payload.url)`. - `defaultmodules/newsfeed/node_helper.js:25-38` performs `fetch(url, { method: "HEAD" })` on the supplied URL and sends the result back. - `defaultmodules/calendar/node_helper.js:13-24` accepts `ADD_CALENDAR`/`FETCH_CALENDAR` socket messages. - `defaultmodules/calendar/node_helper.js:40-58` accepts an arbitrary syntactically valid calendar URL and creates a fetcher. - `defaultmodules/calendar/calendarfetcher.js:35-44` passes the URL into `HTTPFetcher`, whose fetch sink is `js/http_fetcher.js:286-294`. - `defaultmodules/updatenotification/node_helper.js:46-68` accepts `CONFIG`, `MODULES`, and `SCAN_UPDATES` notifications and trusts the socket-provided config/module list. - `defaultmodules/updatenotification/update_helper.js:43-47` stores update commands from config. - `defaultmodules/updatenotification/update_helper.js:96-116` executes the selected update command with `child_process.exec` in the module directory. - `defaultmodules/updatenotification/update_helper.js:221-227` looks up the command from `config.updates` by module name.

False-positive screening performed:

- Express HTTP routes are protected by `ipAccessControl(config.ipWhitelist)` at `js/server.js:89-90`; this does not protect Socket.IO because Socket.IO is attached to the raw HTTP server and no Socket.IO middleware was found. - The explicit `/cors` HTTP endpoint has separate SSRF mitigations (`js/server_functions.js:47-117`) and is disabled by default (`js/defaults.js:14`); the confirmed request primitive here uses module-helper socket paths, not `/cors`. - The command-execution variant is not an unconditional default RCE: `updatenotification` only executes an update command for a non-core git-managed module that is considered behind. However, the trusted command source is attacker-controlled through the unauthenticated socket `CONFIG` message once this boundary is crossed. - Default loopback-only binding lowers default remote exposure, but the sample configuration documents exactly the deployment model where users rely on `ipWhitelist` for network restrictions.

Affected-version evidence: only `magicmirror@2.36.0` at commit `fb41d24ef522e91e802e2a623ff6afbddeb3c9d8` was tested. The affected range is unknown from this audit; earlier versions were not tested. No patched version or fix commit was identified locally.

### PoC The following safe local PoCs were run from a clean checkout of MagicMirror at commit `fb41d24ef522e91e802e2a623ff6afbddeb3c9d8`. Because `node_modules` were not installed in this audit environment and `package.json:52` has a destructive `postinstall` (`git clean -df fonts vendor modules/default`), the commands use small Node harnesses with stubs for missing dependencies while exercising the vulnerable repository code paths directly. They do not contact external hosts and write only disposable `/tmp` marker files.

1. Confirm that the runtime lacks Socket.IO IP allowlist controls:

```bash grep -RIn --exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.claude --exclude-dir=reports -E "allowRequest|io\.use\(|handshake|ipAccessControl\(|cors: \{|origin: /\.\*\$/" js defaultmodules serveronly config tests ```

Observed output:

```text js/server.js:44: cors: { js/server.js:90: app.use(ipAccessControl(config.ipWhitelist)); ```

This confirms the IP allowlist appears only as Express middleware and no Socket.IO handshake/namespace allowlist was present in the reviewed runtime files.

2. Confirm a default module helper will perform a server-side request to an attacker-supplied loopback URL when driven through its socket notification handler:

```bash node -e 'const Module=require("module"); const orig=Module._load; Module._load=(r,p,m)=>{ if(r==="logger") return {log(){},error(){},warn(){},info(){},debug(){}}; if(r==="node_helper") return {create:(o)=>function(){Object.assign(this,o);this.sendSocketNotification=(n,p)=>console.log("SOCKET",n,JSON.stringify(p));}}; if(r==="./newsfeedfetcher") return function(){}; return orig(r,p,m); }; const calls=[]; global.fetch=async(url,opts)=>{calls.push({url,opts}); return {headers:{get:(h)=>h==="x-frame-options"?"deny":null}};}; const Helper=require("./defaultmodules/newsfeed/node_helper"); const h=new Helper(); h.start(); h.socketNotificationReceived("CHECK_ARTICLE_URL",{url:"http://127.0.0.1:65535/internal"}); setTimeout(()=>console.log("fetchCalls=",JSON.stringify(calls)),10);' ```

Observed output:

```text SOCKET ARTICLE_URL_STATUS {"url":"http://127.0.0.1:65535/internal","canFrame":false} fetchCalls= [{"url":"http://127.0.0.1:65535/internal","opts":{"method":"HEAD"}}] ```

Expected vulnerable output: the harness records a server-side `HEAD` request to `http://127.0.0.1:65535/internal` even though `/cors` SSRF protections are not involved.

3. Confirm the conditional command-execution sink is reachable from the trusted socket-driven update path using a harmless `/tmp` marker command:

```bash rm -f /tmp/mm-rce-marker mkdir -p /tmp/mm-audit/modules/evil /tmp/mm-audit/defaultmodules node -e 'const fs=require("node:fs"); const Module=require("module"); const orig=Module._load; Module._load=(r,p,m)=>{ if(r==="logger") return {log(){},error(){},warn(){},info(){},debug(){}}; if(r==="node_helper") return {create:(o)=>function(){Object.assign(this,o);this.sendSocketNotification=()=>{};}}; return orig(r,p,m); }; global.root_path="/tmp/mm-audit"; global.defaultModulesDir="defaultmodules"; fs.writeFileSync("/tmp/mm-audit/defaultmodules/defaultmodules.js","module.exports=[]"); const Helper=require("./defaultmodules/updatenotification/node_helper"); const h=new Helper(); h.gitHelper={add:async()=>{},getRepos:async()=>[{module:"evil",behind:1}],checkUpdates:()=>[{module:"evil",behind:1}]}; h.sendSocketNotification=(n,p)=>console.log("SOCKET",n,JSON.stringify(p)); (async()=>{ await h.socketNotificationReceived("CONFIG",{updates:[{evil:"printf ok > /tmp/mm-rce-marker"}],updateTimeout:5000,updateAutorestart:false,ignoreModules:[],sendUpdatesNotifications:false,updateInterval:60000,useModulesFromConfig:true}); await h.socketNotificationReceived("MODULES",["evil"]); console.log("marker=",fs.readFileSync("/tmp/mm-rce-marker","utf8")); process.exit(0); })();' rm -f /tmp/mm-rce-marker rm -rf /tmp/mm-audit ```

Observed output from the executed harness:

```text SOCKET REPO_STATUS {"module":"evil","behind":1} SOCKET UPDATE_STATUS {"name":"evil","updateCommand":"printf ok > /tmp/mm-rce-marker","inProgress":true,"error":false,"updated":true,"needRestart":true} marker= ok ```

Expected vulnerable output: `marker= ok` demonstrates the update command supplied through the trusted socket configuration path reached `child_process.exec` and wrote the harmless marker.

Negative/control cases:

- With the shipped committed defaults (`js/defaults.js:8-13`), the server binds `localhost` and `ipWhitelist` includes only loopback, so a remote network attacker cannot reach either HTTP or Socket.IO unless the deployment is changed to a documented non-loopback address. - The `updatenotification` command execution path requires at least one non-core module update result; if `checkUpdates()` returns no third-party module with `behind > 0`, `update_helper.parse(...)` does not execute an update command. - The explicit `/cors` route was not used for the confirmed request primitive and has separate protocol/hostname/DNS checks.

Final repro re-check: the sink search, newsfeed socket request harness, and update marker harness were re-run after drafting; the observed outputs above are from this environment. Cleanup removed `/tmp/mm-rce-marker` and `/tmp/mm-audit`.

### Impact In a documented common non-loopback deployment that relies on `ipWhitelist` for access control, an unauthenticated network client can bypass the intended IP allowlist for Socket.IO module namespaces. The attacker can send arbitrary module-helper notifications and payloads as if they were a trusted browser client.

Confirmed impacts include:

- Confidentiality/SSRF: server-side requests to attacker-chosen URLs through default module helpers, including loopback/internal URLs. The newsfeed PoC shows a `HEAD` request to `127.0.0.1`. - Integrity/availability: arbitrary manipulation of module-helper state and periodic fetch/update behavior through trusted socket messages. - Conditional code execution: when a third-party git-managed module is considered behind by the update checker, an attacker can supply a command in the socket `CONFIG` payload and trigger `child_process.exec`; the PoC safely wrote a `/tmp` marker.

CVSS 3.1 rationale for the primary trust-boundary bypass with confirmed SSRF and conditional RCE variant: AV:A because MagicMirror's security policy says it is intended for trusted local/private networks and the documented non-loopback exposure is LAN-style; AC:H because default loopback settings must be changed and the RCE variant requires a pending third-party module update, though SSRF requires fewer conditions after reachability; PR:N because no application authentication is required; UI:N because the attacker connects directly to Socket.IO; S:C because the vulnerable application can cause requests/actions against other local/internal services and can execute commands in a child process in the conditional variant; C:L/I:L/A:L for confirmed internal request and helper-state/command side effects, with conservative scoring because unconditional default RCE was not shown.

### Suggested remediation Apply the same access-control decision to Socket.IO handshakes and namespaces as to Express routes. Concretely, add a Socket.IO `allowRequest` or `io.use(...)` middleware that normalizes `socket.handshake.address`/request IPs with the same `ipAccessControl` logic, and reject clients not allowed by `config.ipWhitelist`. Avoid relying on CORS for authorization; keep origin checks as a browser hardening layer only.

Also add module-helper authorization so arbitrary clients cannot send privileged server-side module notifications. For example, issue an unguessable per-session/module token to the served client and require it on helper messages, or separate read-only client events from privileged server-maintenance actions.

For defense in depth:

- Add SSRF protections or allowlists to calendar/newsfeed/weather helper fetch paths, not only `/cors`. - Do not accept `updatenotification` update commands from socket payloads; use the server-loaded config only, validate module names against configured modules, and avoid shell execution where possible. - Add regression tests proving a disallowed IP receives a rejected Socket.IO handshake even when Express routes are protected, and proving `/newsfeed`/`/calendar` helper messages from unauthenticated sockets cannot trigger server-side requests or update commands.