## Summary The MCP Apps UI resource (an interactive DataStore table rendered in the host's webview/iframe) communicates with its parent via `postMessage`. Its inbound `message` listener performs no `event.origin` validation, and its outbound messages are sent with target origin `'*'`. Any window/frame able to reach the document can inject data into the table and interfere with the MCP UI handshake, and the UI's own JSON-RPC messages are broadcast without an origin restriction. ## Affected code `src/resources/datastore-table-ui.ts` (inline HTML): ```js window.addEventListener('message', function (event) { var msg = event.data; // no check of event.origin if (!msg || typeof msg !== 'object') return; if (msg.id !== undefined && _pending[msg.id]) { ... resolve/reject pending MCP request ... } ... if (data && (data.fields || data.records)) initTable(data); // renders attacker-supplied data }); // outbound: window.parent.postMessage({ jsonrpc: '2.0', id: id, method: method, params: params || {} }, '*'); mcpNot: window.parent.postMessage({ jsonrpc:'2.0', method:method, params:params||{} }, '*'); ``` There is no allowlist of accepted origins for inbound messages and no fixed target origin for outbound messages. ## Impact - **Content spoofing.** A frame that can `postMessage` to the UI can supply arbitrary `fields`/`records`, causing the table to render fabricated rows and links (e.g. a phishing `__link_url`) that appear to come from the queried portal. (Values are HTML-escaped, so this is display spoofing, not script injection — no XSS.) - **Handshake/UX interference.** Because inbound messages matching a pending `id` resolve/reject the UI's MCP requests, an attacker frame can spoof responses to the `ui/initialize` handshake or otherwise disrupt the panel. - **Outbound leakage.** Posting to `'*'` means the UI's JSON-RPC messages are delivered to whatever the parent origin is; if the resource is ever embedded by a context other than the trusted host, those messages leak to it. Exploitability depends on the host embedding model (in a correct MCP Apps host the parent is the trusted host, which limits who can message the frame), so this is a hardening / defense-in-depth issue rather than a directly remotely-triggerable one — hence Low. ## Proof of concept - `poc/verify-no-origin-check.mjs` — static confirmation that the listener has no origin check and outbound uses `'*'`: ``` [x] validates event.origin : NO [x] outgoing postMessage targetOrigin : '*' (any) RESULT: VULNERABLE ``` - `poc/parent-spoof-poc.html` — a page that embeds the UI and injects fabricated rows cross-origin with no handshake; the UI renders them. ## Remediation - Validate `event.origin` against the expected host origin on every inbound message; drop messages from other origins. - Send outbound `postMessage` with an explicit target origin (the host origin established at initialize time) instead of `'*'`. - Optionally, bind to a nonce/channel established during `ui/initialize` so only the negotiated host can drive the panel.